SSTAR: a user-friendly framework for detecting and monitoring subtle thermal precursors to volcanic eruptions—application to Shishaldin, Alaska Authors: Társilo Girona, Laure Brenot Published: 2026 ABSTRACT In recent decades, the monitoring of volcanoes has been revolutionized by the launch of Earth-observing satellites and advances in thermal infrared remote sensing. These developments have revealed a wide range of thermal responses of volcanic surfaces to subsurface processes, even demonstrating that eruptions are often preceded by measurable thermal anomalies. This recognition highlights the need for robust tools to systematically detect and track such anomalies, making full use of existing satellite datasets and maximizing the value of current instruments in orbit. To address this challenge, we present the Subtle Surface Thermal Anomalies Recognizer (SSTAR), a versatile and user-friendly application designed to analyze diffuse thermal anomalies, i.e., subtle thermal unrest (~ 1 K) across large areas (several km2). SSTAR leverages data from NASA’s Terra and Aqua satellites, which host the Moderate Resolution Imaging Spectroradiometers (MODIS), and builds upon a robust statistical framework. By processing pixel-level data, SSTAR tracks the temporal evolution of diffuse thermal anomalies at specific target sites and maps their spatiotemporal distribution across extended areas. Key features include filtering tools that distinguish between long-term (years) and short-term (weeks) anomalies, as well as uncertainty quantification using bootstrapping. The application is standalone, features an interactive interface for streamlined analysis, and is accessible to newcomers to satellite-based thermal remote sensing. At the same time, specialized users can customize the underlying scripts for other specific research needs. As a demonstration, we apply SSTAR to Shishaldin volcano (Alaska), revealing the emergence of significant thermal anomalies around the summit crater and flanks prior to eruptions. We envision SSTAR as a valuable resource for studying subtle thermal unrest at active volcanoes and hydrothermal systems, where the detection of faint and spatially coherent anomalies may help identify subsurface fluid pathways. Its flexible design enables integration with additional satellite datasets, positioning SSTAR as a forward-looking tool for advancing space-based volcanic thermal monitoring. Building on this capability, daily updated diffuse thermal anomalies are provided for target volcanoes through an open web platform hosted at Geosciences Barcelona—CSIC (https://sstar.geo3bcn.csic.es/), to support surveillance agencies and expert committees in alert-level assessments.Graphical Abstract Supplementary InformationThe online version contains supplementary material available at 10.1186/s40623-026-02497-6. Introduction Satellite-based thermal remote sensing provides critical insights into volcanic and hydrothermal activity (Glaze et al. 1989; Francis and Rothery 2000; Vaughan et al. 2012; Haselwimmer and Prakash 2013; Harris 2013; Norini et al. 2015; Pergola et al. 2018; Cigna et al. 2020). It has been instrumental, for example, in detecting eruptions in remote regions (Worden et al. 2014; Girina et al. 2023), characterizing lava flows and lava lakes (Abrams et al. 1991; Oppenheimer and Francis 1997; Wright and Flynn 2003; Coppola et al. 2009, 2010; Patrick et al. 2016; Nádudvari et al. 2020; Thompson and Ramsey 2020; Plank et al. 2023; De Los Reyes et al. 2024), and delineating areas of intense fumarolic discharge or hydrothermally altered soils (Lagios et al. 2007; Mia and Fujimitsu 2013; Vaughan et al. 2020). Over the past three decades, satellite-based thermal remote sensing has also demonstrated considerable potential for detecting volcanic unrest and early eruption precursors (Dehn et al. 2000; Pieri and Abrams 2005; Pergola et al. 2009; van Manen and Dehn 2009; Girina 2012; Murphy et al. 2016; Reath et al. 2016, 2019; Valade et al. 2019; Girona et al. 2021; Gomez-Patron et al. 2025), typically associated with shallow magma emplacement (van Manen and Dehn 2009) or fluid circulation (Laiolo et al. 2017). These studies underscore the importance of improving the detection of pre-eruptive thermal unrest and deepening our understanding of thermal precursors to eruptions. Achieving these goals is essential for strengthening forecasting frameworks and calls for continued efforts to design and refine satellite-based thermal monitoring algorithms. A wide range of algorithms and platforms have been developed to systematically monitor volcanic thermal emissions from space. Early efforts included MODVOLC, which applies a threshold-based NTI (Normalized Thermal Index) to MODIS (Moderate Resolution Imaging Spectroradiometers) data for near–real–time hot-spot detection (Wright et al. 2004); and HOTSAT, which merged MODIS and SEVIRI (Spinning Enhanced Visible and InfraRed Imager) radiances to improve temporal coverage (Ganci et al. 2011). The geostationary system HOTVOLC further exploited SEVIRI’s high-frequency imaging to track anomalies over Europe and Africa (Gouhier et al. 2016). Other contemporaneous systems were AVHotRR, which applies threshold analysis to AVHRR (Advanced Very-High-Resolution Radiometer) channels (Lombardo 2016); RSTVOLC, which uses the Robust Satellite Techniques framework to detect anomalies against long-term reference fields (Lacava et al. 2011; Pergola et al. 2016); VolSatView, a multi-sensor platform (AVHRR, MODIS, VIIRS -Visible Infrared Imaging Radiometer Suite-) for Kamchatka (Gordeev et al. 2016); and MIROVA, which introduced a more sensitive MODIS-based approach by combining MIR (mid-infrared) and TIR (thermal infrared) radiances to compute spectral indices and derive volcanic radiative power (Coppola et al. 2016). More recent efforts include VOLCANOMS, which exploits Landsat imagery for targeted volcano monitoring (Layana et al. 2020); the NHI (Normalized Hotspot Indices) tool, a Google Earth Engine application that integrates data from the Operational Land Imager and the Multispectral Instrument onboard the Sentinel 2 and Landsat 8 satellites, using spectral indices for global-scale detection (Genzano et al. 2020; Mazzeo et al. 2021); the FastVRP platform, which analyzes volcanic radiative power using VIIRS and Sea and Land Surface Temperature Radiometers (SLSTR) (Torrisi et al. 2022); HotLINK, which applies a convolutional neural network to identify thermal anomalies at Alaska volcanoes using VIIRS and MODIS data (Saunders-Shultz et al. 2024); and other machine learning- and deep learning-based algorithms (Corradino et al. 2023; Amato et al. 2023; Cariello et al. 2024). Together, these systems illustrate the transition from simple threshold-based methods to increasingly sophisticated, multi-sensor approaches that enhance the ability to detect volcanic thermal signals. The aforementioned algorithms focus on detecting moderate-to-high temperature features or hotspots, i.e., pixel- or subpixel-scale thermal anomalies that are substantially warmer than ambient conditions and their surroundings, typically linked to fumaroles, lava domes, lava lakes, or lava flows. Building on this foundation, Girona et al. (2021) introduced a different perspective by evaluating the statistical behavior of pixel ensembles rather than relying on threshold- or machine learning-based analyses of individual pixels or scenes. This approach uncovered a previously unreported manifestation of pre-eruptive unrest: very subtle (up to a few kelvins in radiant or brightness temperature), slowly growing (timescales of years) anomalies that extend across large areas (~ km2) of volcanic edifices. These pre-eruptive signals, referred to as diffuse, low-temperature thermal anomalies, reflect subtle changes in background surface temperature or emissivity and are thought to represent the direct or indirect manifestation of magmatic or hydrothermal steam condensing within the shallow subsurface (García-Martínez et al. 2025). Subsequent applications of extended versions of the method to various volcanic settings have shown that diffuse thermal anomalies correlate with independent observations such as seismicity and deformation, supporting its interpretation as a manifestation of subsurface processes and helping to constrain their nature and evolution (Lundgren et al. 2020; Reed et al. 2021; Zhan et al. 2022). Despite these advances, the Girona et al. (2021) method still faces important limitations, making studies of diffuse thermal anomalies highly challenging. First, its applicability is restricted to volcanoes with significant topography due to the strategy used for discarding cloudy scenes. Second, it produces a bulk metric describing only the temporal evolution of the entire edifice, without resolving the spatial distribution of anomalies. And third, the algorithm is computationally intensive, and thus an optimized, user-friendly version is required for practical use worldwide. Addressing these limitations is the primary objective of this study. In this report, we introduce SSTAR -Subtle Surface Thermal Anomalies Recognizer- (Girona and Brenot 2026), a versatile, optimized, user-friendly application designed to retrieve the spatiotemporal distribution of diffuse thermal anomalies at volcanoes or hydrothermal systems while overcoming the main limitations of the Girona et al. (2021) algorithm. The paper is organized as follows: Section 2 outlines the fundamentals of the algorithm and highlights the improvements introduced in SSTAR relative to the approach of Girona et al. (2021) for analyzing diffuse thermal anomalies at volcanoes. Section 3 provides a detailed description of the application and methodological workflow. Section 4 presents the results obtained for Shishaldin volcano (Alaska) as an illustrative case study. Section 5 discusses these results, places diffuse thermal anomalies in context with high-temperature thermal anomalies, and outlines recommended best practices for the reliable application of SSTAR. Finally, Section 6 summarizes the main conclusions. Fundamentals of the SSTAR algorithm The SSTAR algorithm is implemented in MATLAB and builds upon the automated median-of-medians framework developed by Girona et al. (2021) for analyzing satellite-derived thermal infrared data from MODIS sensors. Based on this earlier framework, SSTAR introduces several refinements aimed at improving data consistency, minimizing noise, and enhancing the spatial characterization of thermal anomalies. These include an optimized cloud-filtering procedure, adaptive spatial aggregation that captures anomalies independently of topographic effects, and a set of filtering options that increase the signal-to-noise ratio while preserving data trends. Together, these developments allow the applicability of SSTAR to a wide range of volcanic settings and enable an accurate assessment of surface thermal activity over time. A detailed comparison between SSTAR and the Girona et al. (2021) algorithm is provided in Table 1. Table 1Comparison between the SSTAR algorithm (this work) and the approach presented by Girona et al. (2021), highlighting the main methodological and functional differences between both implementationsAlgorithmSSTAR (this work)Girona et al. (2021) approachUsabilityUser-friendly applicationNo user-friendly interfaceCloud filteringBased on cloud mask product(discards cloudy pixels)Based on topographic effects on temperature (discards entire scenes)ApplicabilityAny magmatic systemSystems with significant topographyRegion of interestWell defined and user selectedNot well defined; depends on pixels with the lowest brightness temperature, which vary with timeReference areaWell defined and user selectedNot well defined; depends on pixels with the highest brightness temperature, which vary with timeTemporal filteringThe user can select multiple options Only one option: Iterative wavelet denoising followed by a low-pass, backward-looking median filterFundamental metric trackedMedian brightness temperature of a region of interest (spatially fixed over time)Median brightness temperature of the coolest area of the volcanic edifice (slightly variable over time)Type of resultsSpatiotemporal distribution of anomalies (full control over the location of anomalies)Temporal evolution of anomalies (no spatial information on anomalies) Table 1Comparison between the SSTAR algorithm (this work) and the approach presented by Girona et al. (2021), highlighting the main methodological and functional differences between both implementationsAlgorithmSSTAR (this work)Girona et al. (2021) approachUsabilityUser-friendly applicationNo user-friendly interfaceCloud filteringBased on cloud mask product(discards cloudy pixels)Based on topographic effects on temperature (discards entire scenes)ApplicabilityAny magmatic systemSystems with significant topographyRegion of interestWell defined and user selectedNot well defined; depends on pixels with the lowest brightness temperature, which vary with timeReference areaWell defined and user selectedNot well defined; depends on pixels with the highest brightness temperature, which vary with timeTemporal filteringThe user can select multiple options Only one option: Iterative wavelet denoising followed by a low-pass, backward-looking median filterFundamental metric trackedMedian brightness temperature of a region of interest (spatially fixed over time)Median brightness temperature of the coolest area of the volcanic edifice (slightly variable over time)Type of resultsSpatiotemporal distribution of anomalies (full control over the location of anomalies)Temporal evolution of anomalies (no spatial information on anomalies) Comparison between the SSTAR algorithm (this work) and the approach presented by Girona et al. (2021), highlighting the main methodological and functional differences between both implementations Based on cloud mask product (discards cloudy pixels) Specifically, SSTAR transforms multiyear thermal infrared observations into interpretable spatiotemporal records of surface thermal anomalies. It integrates spectral radiance, geolocation, and cloud-mask datasets acquired from MODIS instruments, which can be retrieved from NASA’s Earthdata repository (https://earthdata.nasa.gov/). The algorithm uses the MOD021KM/MYD021KM products (Terra/Aqua Calibrated Radiances, 1 km, Level 1B, Collection 6), which provide radiometrically calibrated radiances corrected for instrumental effects, whereas pixel coordinates are derived from the MOD03/MYD03 products (Terra/Aqua Geolocation Fields, Level 1A) and cloud-free pixels are identified using the MOD35/MYD35 products (Terra/Aqua Cloud Mask, Level 2, Collection 6.1). For each valid scene, spectral radiance values from band 31 (10.78–11.28 μm) are extracted from cloud-free pixels. From the preprocessed satellite data, the median spectral radiance is first computed for each region of interest and for a reference area (Fig. 1), and then converted to median brightness temperatures using Planck’s law (e.g., Blackett 2017). These values are then subtracted and aggregated into a daily median, producing a regularly sampled time series with one value per day that provides a continuous record of relative thermal variations. To ensure temporal continuity, missing days are filled using predefined interpolation schemes. Because the resulting series contains both seasonal fluctuations and random variability, several filtering techniques are available to isolate the long-term trends characteristic of diffuse thermal anomalies. The filtered output, referred to as the diffuse thermal anomaly time series, captures these long-term variations in median brightness temperature relative to the reference area. The diffuse thermal anomaly time series are subsequently analyzed to quantify uncertainties using a statistical method that balances robustness with computational efficiency. Specifically, we apply a bootstrap resampling approach that creates synthetic datasets by randomly resampling the original observations with replacement, enabling confidence intervals to be estimated without assuming any particular underlying distribution. This method is both efficient and well-suited for characterizing the variability inherent in the median-based statistics employed by SSTAR.Fig. 1Schematic illustration of the core elements of the SSTAR algorithm. The red points mark the user-selected target locations, the green area denotes the reference region, and the blue blocks represent the regions of interest (ROIs), which are outlined by dark-blue boundaries. Grey squares indicate cloud-contaminated pixels, which are excluded from the analysis. For each ROI, the statistical calculations incorporate all cloud-free (non-grey) pixels within both the ROI and the reference area Schematic illustration of the core elements of the SSTAR algorithm. The red points mark the user-selected target locations, the green area denotes the reference region, and the blue blocks represent the regions of interest (ROIs), which are outlined by dark-blue boundaries. Grey squares indicate cloud-contaminated pixels, which are excluded from the analysis. For each ROI, the statistical calculations incorporate all cloud-free (non-grey) pixels within both the ROI and the reference area Once the median anomaly time series and its uncertainty have been computed for all regions of interest, SSTAR provides multiple options for visualization and interpretation. Users can display individual time series to examine the temporal evolution of surface thermal activity at specific locations, identifying long-term trends, transient fluctuations, or potential precursory signals. Alternatively, multiple time series can be combined to generate spatially distributed representations of diffuse thermal anomalies through time. These visualizations, presented as spatiotemporal maps or animations, reveal how thermal patterns propagate, migrate, or persist across the study area, offering an integrated perspective on both the temporal and spatial dynamics of surface thermal infrared anomalies. Step-by-step description of the user-friendly SSTAR application The SSTAR application provides an end-to-end workflow that transforms raw multispectral MODIS swath data into scientifically ready time-series and spatial products of diffuse thermal anomalies of the Earth’s surface. The application integrates all stages of analysis, from data preprocessing and signal extraction to uncertainty quantification and visualization, within a single, user-friendly interface (Fig. 2). Designed for accessibility, SSTAR enables users with limited experience in programming, satellite data handling, or thermal infrared analysis to conduct sophisticated thermal investigations that would otherwise require extensive manual scripting.Fig. 2SSTAR interface. The application offers an interactive platform for visualizing and analyzing diffuse thermal anomalies at active volcanoes and hydrothermal systems. Through its different tabs (Data Processing, Time Series, and Spatiotemporal Map), users can define parameters such as the region of interest, reference area, target locations, filtering method, and interpolation scheme. Results are updated automatically, enabling rapid and efficient exploration of thermal trends and sensitivity analyses. The main display area (indicated by the logo) presents key outputs, including the analysis area, target locations, and the spatiotemporal evolution of thermal anomalies SSTAR interface. The application offers an interactive platform for visualizing and analyzing diffuse thermal anomalies at active volcanoes and hydrothermal systems. Through its different tabs (Data Processing, Time Series, and Spatiotemporal Map), users can define parameters such as the region of interest, reference area, target locations, filtering method, and interpolation scheme. Results are updated automatically, enabling rapid and efficient exploration of thermal trends and sensitivity analyses. The main display area (indicated by the logo) presents key outputs, including the analysis area, target locations, and the spatiotemporal evolution of thermal anomalies At its core, the software automates the extraction, filtering, processing, and normalization of spectral radiance data from MODIS band 31, sensitive to surface thermal emissions (Ahmad et al. 2002); handles geolocation and cloud masking, and outputs consistent temporal and spatial datasets suitable for scientific interpretation. Users can define regions of interest (ROIs), reference zones, and analysis parameters interactively through the graphical interface, with built-in diagnostics that facilitate error checking and reproducibility. Additionally, SSTAR integrates a suite of interactive visualization tools, allowing users to zoom, pan, or adjust axes for closer data inspection of specific time windows. The visualization interface allows users to explore processed outputs directly within the software environment, ensuring a seamless transition between computation and analysis. A save option automatically exports each figure as a sequentially numbered PNG file, ensuring an organized record of visualization outputs. Meanwhile, a real-time log window documents the analysis steps, parameter settings, and progress messages, which facilitates iterative testing and diagnostics. Together, these features provide a coherent and efficient framework for reviewing results, verifying data integrity, and correlating satellite-derived thermal signals with independent records of volcanic activity. Below, we present the SSTAR application using Shishaldin volcano (Alaska) as an illustrative case, outlining the logical sequence of operations and the key user decisions at each stage, from organizing MODIS input files to generating thermal anomaly time series and maps. These sections emphasize the conceptual and methodological workflow rather than specific software instructions; a detailed user´s guide is provided in the HELP menu of the application. Data processing: organization and initialization After the MODIS data have been downloaded, the complementary datasets (calibrated radiances, geolocation fields, and cloud-mask products) must be stored in separate directories for Terra and Aqua, following a consistent hierarchical structure. When the user initializes a new project, SSTAR scans these directories, verifies data completeness, and associates each radiance file with its corresponding geolocation and cloud-mask counterparts. The user only needs to specify the main data folder and an output directory to store the results, after which the application automatically builds the analysis environment. All computations in SSTAR are performed with respect to a fixed geographic reference that defines the origin of the reference frame for the analysis grid and serves as the basis for distance calculations and map projections. In the case of Shishaldin volcano, this reference is fixed to the summit coordinates 54.756° N and 163.97° W, as retrieved from the Global Volcanism Program database (Andrews et al. 2025). Using this fixed origin ensures that all calculations and visualizations are expressed relative to the edifice center. Before proceeding with the analysis, SSTAR allows users to visualize the defined area on a satellite basemap, providing an additional check that the spatial origin is correctly positioned. Once the data environment has been initialized and verified, the workflow proceeds to the definition of the target locations and subsequent processing steps. Data processing: target locations, completeness, and extraction of descriptive metrics Once the data environment is initialized, the next step is to specify the target locations of the study. The target locations can be defined in two ways. The first employs a structured distribution of target locations, specified by western/eastern and southern/northern limits and by a uniform spatial step in kilometers, which is suited for systematic mapping of the thermal anomaly field around a volcano. The second allows the selection of specific target locations at selected distances from the reference location, useful for targeted analyses such as crater-focused or sectoral studies. In both cases, SSTAR can overlay the selected domain on a satellite basemap, allowing for visual verification that the area of analysis is correctly positioned (Fig. 3a). Once this is set, the preprocessing step of the application is ready to be run.Fig. 3Visualizations provided by SSTAR. a Map view showing the spatial configuration of target locations used for thermal analysis. Each target location represents an independent unit where brightness temperature time series are computed for different regions of interest. b Example of a diagnostic color-coded map generated by the application, showing the spatial distribution of analysis completeness. Colors represent the relative degree of completeness, offering a quick and intuitive overview of regions that may require reprocessing or additional data runs. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) Visualizations provided by SSTAR. a Map view showing the spatial configuration of target locations used for thermal analysis. Each target location represents an independent unit where brightness temperature time series are computed for different regions of interest. b Example of a diagnostic color-coded map generated by the application, showing the spatial distribution of analysis completeness. Colors represent the relative degree of completeness, offering a quick and intuitive overview of regions that may require reprocessing or additional data runs. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) A key feature of SSTAR is its ability to maintain analytical continuity and ensure reproducibility throughout the processing workflow. Given the large size of satellite archives and the computational intensity of multi-year analyses, interruptions during processing—whether due to hardware limitations, software updates, or user intervention—are sometimes inevitable. To address this, SSTAR implements a checkpoint-based structure in which progress is stored independently for each grid cell and satellite sensor. This modular organization allows the software to resume from the point of interruption without repeating completed calculations, thereby preserving computational efficiency and minimizing data loss. When an analysis session is reopened, the user can inspect a completion plot that summarizes which target locations were successfully processed. This diagnostic step calculates the percentage of satellite swaths successfully processed for each target location and displays the results as a color-coded map (Fig. 3b). By highlighting areas with incomplete coverage, SSTAR allows users to re-run only the missing or incomplete sub-sectors, seamlessly integrating the new results into the existing dataset. The software also includes safeguards to detect and repair partially written text files, truncating corrupted outputs when it is safe to do so. This automated error-handling mechanism contributes to the stability of long-term analyses and reduces the need for manual data management. Upon initiation of the analysis, SSTAR applies the MOD35/MYD35 cloud mask to filter out cloud-contaminated pixels, retaining those classified as “confident clear” and excluding pixels flagged as “probably clear,” “probably cloudy,” or “cloudy.” This conservative approach minimizes contamination from thin clouds, cirrus, and mixed pixels and, although it may reduce data availability, improves the robustness required to reliably detect subtle, low-amplitude thermal anomalies. After cloud filtering, SSTAR computes representative spectral radiance values for each target location. One of the representative values is the median spectral radiance of all cloud-free pixels within a square area of side x centered on the target location. Another representative value is the median spectral radiance of the surrounding background zone, defined as the region between the outer edge of the inner square (side x) and a larger concentric square with a side length of 80 km. This process is repeated for all selected target locations and for up to 44 distinct square sides, covering spatial scales from small regions of x = 2 km per side (4 km2) to broad areas up to x = 70 km per side (4900 km2). By systematically varying the spatial scale, SSTAR characterizes the surface multiscale thermal behavior, allowing for exploration of anomalies associated with both localized volcanic features and regional-scale thermal patterns related to broader geophysical or climatic processes. Additionally, SSTAR records several diagnostic parameters for each target location, including the scene date, the number of cloud-free pixels used in each median spectral radiance estimate, and the illumination conditions (i.e., daytime or nighttime). The extracted thermal values along with their corresponding descriptive metrics are saved in one output file per target location and sensor. To guarantee full reproducibility, the software also generates a configuration log (config.txt) at the start of each run that consolidates all relevant information, including input directories, parameter settings, and processing options. This file provides a comprehensive record of the analysis setup, ensuring traceability, reproducibility, and the possibility to replicate or extend the computation at any subsequent stage. The approach promotes transparent data handling and supports rigorous scientific reporting of results derived from the application. It is also worth noting that SSTAR is built with a modular architecture that allows independent groups of target locations to be processed simultaneously. By distributing computations across multiple processing units, this structure significantly reduces computation time, particularly when analyzing large volcanic regions or multidecade satellite archives, while preserving the integrity and internal consistency of the dataset. Construction of the diffuse thermal anomaly time series Once the initial processing step is complete, the SSTAR workflow transitions from data preparation to signal extraction and time-series analysis. At this stage, the previously defined target locations are used to automatically delineate both the regions of interest (ROI) and a reference area. Specifically, the user selects the ROI and reference side lengths from a predefined set of options corresponding to the parameter x used during the initial processing step. After the ROI and reference areas are established, SSTAR offers visualization tools that enable the user to review the selected regions and confirm the correctness of the analysis setup before proceeding (Fig. 4). The user may also choose to analyze a single target location in detail or process all target locations for which representative and diagnostic parameters were successfully retrieved. Then, the user defines the parameters that govern the construction of the thermal time series. Several options are available to tailor the analysis to different research objectives; these include the selection of light conditions (daytime, nighttime, or both), the interpolation method used to fill temporal gaps, and the filtering strategy applied to extract temporal trends of interest.Fig. 4Example of the target locations, regions of interest (ROIs), and reference area selected within the SSTAR framework for analyzing diffuse thermal anomalies at Shishaldin volcano. Satellite map showing the ROI (blue), reference area (green), and all target locations analyzed (red points). Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) Example of the target locations, regions of interest (ROIs), and reference area selected within the SSTAR framework for analyzing diffuse thermal anomalies at Shishaldin volcano. Satellite map showing the ROI (blue), reference area (green), and all target locations analyzed (red points). Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) Available interpolation methods include the nearest-neighbor, linear, spline, piecewise cubic Hermite (PCHIP), and modified Akima cubic Hermite (Makima) schemes (e.g., Teegavarapu 2024). The nearest-neighbor interpolation preserves the original observations and avoids introducing artificial smoothing, making it useful when data are sparse or when retaining the exact measured values is desirable; however, this interpolation scheme produces a step-like signal with abrupt transitions between successive data points, potentially resulting in artificial discontinuities and a less realistic temporal evolution. The linear interpolation generates straight-line transitions between adjacent observations, providing a simple and computationally efficient representation of temporal variations; while it avoids the discontinuities associated with the nearest-neighbor approach, it may underestimate the curvature of the underlying signal and attenuate sharp thermal peaks. The spline interpolation yields a smoother curve that captures the overall temporal trends but can slightly overestimate values across large data gaps, occasionally introducing small oscillations before and after sharp changes. The PCHIP interpolation offers a compromise between smoothness and local accuracy, avoiding unrealistic oscillations while preserving both the natural shape of the signal and the amplitude of the main thermal peaks; compared with spline-based approaches, however, the resulting curve may exhibit slightly sharper transitions between successive observations. The Makima interpolation behaves similarly to PCHIP but introduces slightly smoother transitions, at the expense of a modest reduction in peak amplitudes. For temporal filtering, SSTAR offers two main strategies: (i) a non-causal approach that applies iterative wavelet denoising followed by a low-pass, backward-looking median filter (as introduced in Girona et al. 2021), and (ii) a fully causal approach that uses only a low-pass, backward-looking median filter to isolate multi-year trends. For completeness, SSTAR also allows the use of (iii) a Butterworth high-pass filter to highlight short-term fluctuations, and (iv) a Butterworth band-pass filter to isolate specific frequency ranges. These options allow users to tailor the filtering strategy to the temporal scale of the process of interest, although approaches (i) and (ii), typically implemented with 1-year or 2-year cutoff periods, are the most suitable for detecting the long-term thermal trends characteristic of diffuse thermal anomalies. It is worth noting that strategy (i) effectively suppresses noise and highlights long-term (multi-year) trends, although it may introduce a temporal offset. Alternatively, strategy (ii) can be used to avoid a potential time shift, albeit with slightly higher residual noise. Moreover, the user can also define a minimum acceptable fraction of clear (cloud-free) pixels per ROI, referred to as the acceptance threshold, so that only high-quality scenes are retained for analysis. The use of the acceptance threshold makes sense when using large enough ROIs, with a significant number of pixels within the ROI. Finally, the user can control the quality of uncertainty quantification by specifying the number of realizations used to compute the 95% confidence envelope around the median signal. This step is implemented through a bootstrapping procedure following the approach of Girona et al. (2021). In practice, a few hundred realizations are typically sufficient to obtain stable and reliable uncertainty bounds. The flexible parameter configuration enables users to balance data completeness with temporal smoothing, resulting in optimized time series that can effectively capture both short-term variability and long-term thermal trends, along with their associated uncertainties. Upon completion of the execution, SSTAR generates a [scenario]_results.txt file containing the observation dates, thermal anomaly values (defined as the filtered difference between the median brightness temperature of the ROI and that of the reference area), and their corresponding 95% confidence limits. Additional diagnostic information, such as data gaps, interpolation performance, and scene coverage, is summarized in a [scenario]_summary_analysis.txt file. This structure provides users with both the quantitative results and the necessary metadata to assess the reliability and completeness of the analysis, facilitating subsequent interpretation and integration with complementary datasets. The resulting thermal anomaly time series for each ROI represents relative temporal variations with respect to the reference area. For simplicity and visualization purposes, the series is normalized by subtracting its minimum value (min-shift), ensuring that all entries remain positive without altering the relative variations or trends in the data. Building on this foundation, SSTAR also enables visualization of the independent evolution of the interpolated and filtered daily median brightness temperatures for both the ROI and the reference area, without applying the subtraction. This additional visualization helps assess whether certain trends in the anomaly plots may result from residual climatic or environmental effects that were not fully removed during the differencing and filtering procedures (see more details in Section 5.4.5). To enhance temporal accuracy, SSTAR also offers an optional correction that compensates for potential temporal lags introduced by the iterative wavelet and median filters used to extract multi-year trends (Girona et al. 2021). This correction, defined in days, can be calibrated against independently dated events (such as eruptions, wildfires, or other independently known events), ensuring that the timing of filtered anomalies aligns with real-world phenomena. This correction is not needed when using other predefined filters. Finally, SSTAR facilitates contextual interpretation by allowing external event overlays. By importing a simple reference_dates.csv file listing known eruptions, seismic crises, wildfires that may affect the interpretation, or other relevant events, users can add vertical markers to the time-series plots. These markers provide a temporal framework for comparing thermal fluctuations with independently known activity, helping to determine whether observed anomalies precede, coincide with, or follow documented events. This capability enables a direct assessment of the potential forecasting value of the derived thermal signals. Spatiotemporal mapping and animations After analyzing the regions of interest for all the predefined target locations, the analysis naturally extends from the temporal to the spatial domain, allowing users to explore how thermal anomalies evolve across the volcanic edifice and its surroundings. This step transforms a collection of point-based measurements into spatially continuous maps and animations, providing a synoptic view of the system’s thermal behavior through time and space. Mapping in SSTAR is based on precomputed outputs generated during earlier processing stages. The user first selects a scenario to visualize, corresponding to a specific configuration of filters and interpolation settings, and then defines a temporal reference period that serves as the baseline for spatial anomaly calculations. Diffuse thermal anomaly maps are subsequently expressed relative to the thermal conditions observed during this user-selected time interval, thereby highlighting departures from a representative background state. This normalization enables consistent comparisons between different epochs, especially when analyzing long-term trends spanning decades of satellite observations. The application offers several rendering options that emphasize different aspects of the spatial variability. Maps can be displayed as absolute fields, showing the raw magnitude of the thermal signal; as spatially normalized fields, highlighting local deviations relative to surrounding values; or as normalized cumulative fields, which integrate the thermal evolution over time to reveal the most common patterns. Color scales, transparency levels, and grid density can all be customized interactively, allowing users to fine-tune visual contrast and readability according to the specific volcano morphology or study objective. To improve interpretability, SSTAR allows users to incorporate ancillary information directly into the maps. For instance, when wildfire detections are available, these can be excluded from the normalization and visualization through a simple wildfires.csv file located in the output directory, preventing misidentification of wildfires or other non-volcanic thermal sources (e.g., anthropogenic). Similarly, temporal corrections (which may be needed when using the iterative wavelet approach combined with a median low-pass filter) derived during time-series analysis can also be applied to the spatiotemporal map to adjust the phase of long-term signals, aligning mapped anomalies more precisely with independent events. Once the spatial representation is defined, SSTAR can export the results either as single-date figures for static inspection or as complete spatiotemporal animations in AVI format. In addition, the spatiotemporal results can be exported as CSV files, allowing users to perform further quantitative analysis or integrate the results with complementary datasets in external software. These spatiotemporal results provide a powerful means of visualizing the dynamic progression of thermal anomalies, revealing propagation patterns, migration trends, or localized pulses of activity that may not be evident from individual time series alone. By combining temporal continuity with spatial context, the mapping and animation tools in SSTAR serve as an essential bridge between data processing and scientific interpretation, transforming raw satellite observations into spatiotemporally coherent visualizations of volcanic thermal evolution. Results: Shishaldin volcano (Alaska) as a case example We present below the results obtained from exploring different configuration options available in SSTAR. This analysis is intended not only to demonstrate the capabilities of the method but also to establish a systematic workflow that we recommend applying when using SSTAR for a new volcano, dataset, or time period, thereby ensuring the robustness and reproducibility of the results by evaluating how methodological choices influence the detected thermal signal. In the following sections, we use as a reference area an annular square extending from 60 to 80 km from the volcano center, identical to the configuration shown in Fig. 4. This configuration serves as a baseline throughout the analysis. We then independently vary specific parameters or processing steps to assess their individual effects on the retrieved thermal anomaly time series and maps. The influence of the reference area is further discussed in the Discussion section. Temporal evolution of thermal anomalies at the summit area In this subsection, we present the temporal evolution of the low-temperature thermal anomaly (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\delta T }$$\end{document}δT¯; in K) for two different spatial configurations of the region of interest (ROI) (Fig. 5). In the first configuration (Fig. 5a), corresponding to an ROI of 4 km2 (2 km per side), the thermal anomaly displays marked variability, with sharp peaks coinciding or following eruptive episodes, and reaching values above 3 K during 2020–2021 and again around 2023–2024. Between eruptions, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\delta T }$$\end{document}δT¯ generally remains at low values, reflecting background thermal conditions. The high sensitivity of this smaller ROI allows for the detection of localized and transient brightness temperature increases near the vent area. In contrast, the second configuration (Fig. 5b), corresponding to an ROI of 16 km2 (4 km per side), yields a smoother anomaly curve with lower amplitude fluctuations. Peak values during eruptive periods remain evident but are slightly attenuated relative to the 4 km2 case. The broader averaging over a larger area reduces short-term variability while preserving the overall trend of rising \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\delta T }$$\end{document}δT¯ prior to and during eruptions. The choice of the ROI size affects both the magnitude and temporal character of the signal: smaller ROIs are more sensitive to localized thermal sources, while larger ROIs capture broader, spatially integrated thermal signals.Fig. 5Results obtained with SSTAR for Shishaldin volcano using different regions of interest centered on the summit. a Results obtained for an ROI centered at coordinates (0, 0) with a 2 km side length. b Results obtained for an ROI centered at coordinates (0, 0) with a 4 km side length. In both cases, we use data acquired under both daytime and nighttime conditions, linear interpolation, low-pass (iterative wavelet + median) filtering with a 365-day cutoff period, acceptance threshold of 0%, 100 iterations for uncertainty quantification, and a temporal correction of 180 days (shifting the time series forward). The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated uncertainty at 95% confidence. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) Results obtained with SSTAR for Shishaldin volcano using different regions of interest centered on the summit. a Results obtained for an ROI centered at coordinates (0, 0) with a 2 km side length. b Results obtained for an ROI centered at coordinates (0, 0) with a 4 km side length. In both cases, we use data acquired under both daytime and nighttime conditions, linear interpolation, low-pass (iterative wavelet + median) filtering with a 365-day cutoff period, acceptance threshold of 0%, 100 iterations for uncertainty quantification, and a temporal correction of 180 days (shifting the time series forward). The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated uncertainty at 95% confidence. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) Both spatial configurations show a consistent pre-eruptive increase in thermal anomaly preceding the 2014–2015, 2019–2020, and 2023 eruptive episodes. In particular, the 2014–2015 episode was preceded by a ~ 1 K increase in ~ 4 years for both configurations. The 2019–2020 episode was preceded by a ~ 0.5 K increase in ~ 3 years for both configurations, although the 4 km2 case does not show a monotonous increase. The 2023 eruption was preceded by a ~ 1.5 K and ~ 1 K increase in ~ 2 years for the 4 km2 and the 16 km2 configuration, respectively. The 16 km2 configuration also exhibits a smaller thermal increase prior to the 2008 event (considered unconfirmed; Andrews et al. 2025), although this signal is absent in the 4 km2 configuration. The pre-eruptive anomaly associated with the 2004 episode cannot be evaluated because it happened at the beginning of the time series. Influence of the incident light conditions The thermal anomaly fields derived from the daytime dataset and the nighttime dataset (Fig. 6) show distinct expressions of thermal activity at Shishaldin´s summit (compare with Fig. 5b). In the daytime series (Fig. 6a), the 2014–2016, 2019–2020, and 2023 eruptions are all preceded by a gradual intensification of summit-centered anomalies, although the pre-eruptive summit signal preceding the 2019–2020 sequence remains very weak (~ 0.3 K vs. ~ 1–1.5 K).Fig. 6Results obtained with SSTAR for Shishaldin volcano under different incident light conditions. a Results obtained from data acquired under daytime conditions. b Results obtained from data acquired during nighttime conditions. In both cases, we use an ROI centered at coordinates (0, 0) with a 4 km side length, linear interpolation, low-pass (iterative wavelet + median) filtering with a 365-day cutoff period, acceptance threshold of 0%, 100 iterations for uncertainty quantification, and a temporal correction of 180 days (shifting the time series forward). The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) Results obtained with SSTAR for Shishaldin volcano under different incident light conditions. a Results obtained from data acquired under daytime conditions. b Results obtained from data acquired during nighttime conditions. In both cases, we use an ROI centered at coordinates (0, 0) with a 4 km side length, linear interpolation, low-pass (iterative wavelet + median) filtering with a 365-day cutoff period, acceptance threshold of 0%, 100 iterations for uncertainty quantification, and a temporal correction of 180 days (shifting the time series forward). The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) The nighttime data also reveal additional thermal features (Fig. 6b): pre-eruptive anomalies are not detected at the summit prior to the 2019–2020 eruption, whereas the 2014–2016 and 2023 eruptions were preceded by anomalies of approximately 0.6 K and 1 K, respectively. In addition, the nighttime dataset captures signals that are absent from the daytime record. Most notably, the small 2008 event, which occurred in winter, was preceded by a robust and well-defined ~ 1 K increase in the nighttime sequence, while no corresponding feature is visible in the daytime sequence. Influence of the interpolation method To assess the impact of interpolation methods on gap-filling, below we show the temporal evolution of the low-temperature thermal anomaly (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\delta T }$$\end{document}δT¯) using the different interpolation schemes provided by SSTAR (Fig. 7), allowing direct comparison with the results from linear interpolation (Fig. 5). Overall, all interpolation schemes reproduce similar primary pre-eruptive increases and eruptive peaks in the Shishaldin thermal anomaly record, capturing both the timing and general shape of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\delta T }$$\end{document}δT¯ variations when using a 16 km2 ROI (Fig. 5b). In particular, the nearest-neighbor interpolation scheme, although a less realistic approach to fill data gaps, reproduces the main eruptive peaks and pre-eruptive trends (Fig. 7a). The spline algorithm, however, tends to produce slightly higher anomaly amplitudes—by about 1 K—before the 2014–2016 eruptive episodes (Fig. 7b), likely due to its tendency to oversmooth and extend curvature across data gaps, which can artificially enhance values around steep gradients. In contrast, the PCHIP and Makima algorithms more faithfully preserve the natural evolution of the signal, maintaining both its amplitude and phase without introducing spurious oscillations or distortions (Fig. 7c, d). Altogether, the results indicate that even in a case like Shishaldin, where cloud cover in the Aleutias leads to data gaps of up to ~ 40%, all interpolation methods successfully recover the dominant thermal trends when using an ROI of 16 km2, achieving a good balance between smoothness and fidelity to the underlying data. Differences between schemes become more pronounced as data gaps increase and ROI size decreases, leading to less reliable trend estimates.Fig. 7Results obtained with SSTAR for Shishaldin volcano using different interpolation schemes. a Results obtained using nearest-neighbor interpolation. b Results obtained using spline interpolation. c Results obtained using piecewise cubic Hermite (PCHIP) interpolation. d Results obtained using modified Akima cubic Hermite (Makima) interpolation. In all the cases, we use an ROI centered at coordinates (0, 0) with a 4 km side length, data acquired under both daytime and nighttime conditions, low-pass (iterative wavelet + median) filtering with a 365-day cutoff period, acceptance threshold of 0%, 100 iterations for uncertainty quantification, and a temporal correction of 180 days (shifting the time series forward). The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) Results obtained with SSTAR for Shishaldin volcano using different interpolation schemes. a Results obtained using nearest-neighbor interpolation. b Results obtained using spline interpolation. c Results obtained using piecewise cubic Hermite (PCHIP) interpolation. d Results obtained using modified Akima cubic Hermite (Makima) interpolation. In all the cases, we use an ROI centered at coordinates (0, 0) with a 4 km side length, data acquired under both daytime and nighttime conditions, low-pass (iterative wavelet + median) filtering with a 365-day cutoff period, acceptance threshold of 0%, 100 iterations for uncertainty quantification, and a temporal correction of 180 days (shifting the time series forward). The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) Influence of the filter type In this subsection, we compare the temporal evolution of the low-temperature thermal anomaly (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\delta T }$$\end{document}δT¯) obtained using two distinct denoising approaches: the iterative wavelet combined with a 2-year backward-looking median filter (Fig. 8a) and the 2-year backward-looking median filter alone (Fig. 8b). The time series processed with the iterative wavelet and median filtering exhibits a smooth and coherent long-term signal that effectively suppresses short-term noise while preserving the main thermal trends and pre-eruptive increases. The efficiency of this denoising technique to retrieve hidden trends in noisy data was previously demonstrated in Girona et al. (2021). The resulting anomaly curve shows gradual increases in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\delta T }$$\end{document}δT¯ preceding all eruptive episodes, with well-defined peaks associated with the 2008, 2014–2016, and 2023 events, and consistently lower, stable values during intervals with lower activity. It is worth noting that the iterative wavelet method combined with a 2-year median filter also reveals a diffuse, progressive anomaly within the 16 km2 summit-centered area prior to the 2019–2020 eruption. However, this pre-eruptive signal is considerably weaker than the other anomalies in the time series and does not reach significance at the 95% confidence level.Fig. 8Results obtained with SSTAR for Shishaldin volcano using different filtering schemes. a Low-pass iterative wavelet + median (730-day cutoff period—backward window) filter. Temporal correction of 180 days (moving time series to the future) is applied. b Low-pass median (730-day cutoff period—backward window) filter. In both cases, we use an ROI centered at coordinates (0, 0) with a 4 km side length, linear interpolation, data acquired under both daytime and nighttime conditions, acceptance threshold of 0%, and 100 iterations for uncertainty quantification. The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) Results obtained with SSTAR for Shishaldin volcano using different filtering schemes. a Low-pass iterative wavelet + median (730-day cutoff period—backward window) filter. Temporal correction of 180 days (moving time series to the future) is applied. b Low-pass median (730-day cutoff period—backward window) filter. In both cases, we use an ROI centered at coordinates (0, 0) with a 4 km side length, linear interpolation, data acquired under both daytime and nighttime conditions, acceptance threshold of 0%, and 100 iterations for uncertainty quantification. The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) The time series derived using only the 2-year median filter (Fig. 8b) retains a higher level of high-frequency variability, resulting in a noisier signal with numerous short-lived fluctuations superimposed on the broader thermal trend. This short-term variability partly obscures long-term patterns, although the overall increases in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\delta T }$$\end{document}δT¯ preceding eruptive events remain clearly discernible. Interestingly, this filter configuration provides a more significant upward trend of ~ 1 K prior to the unconfirmed 2008 event and the 2019–2020 eruption. Overall, this comparison shows that the iterative wavelet-based denoising step substantially enhances the signal-to-noise ratio of the thermal anomaly record, yielding a cleaner time series that highlights long-term variations associated with volcanic unrest. The 2-year backward median filter, however, remains a very valuable complementary tool. Although it produces a lower signal-to-noise ratio, it provides a conservative and fully independent reference that ensures the trends identified with the wavelet approach are not artifacts of temporal alignment or filtering choices (Girona et al. 2021). An additional advantage is that it avoids potential temporal shifts, reducing user concerns about distortions introduced during smoothing. Notably, several pre-eruptive patterns at Shishaldin appear more clearly with the backward median filter, and their consistency across both techniques strengthens their interpretation as genuine precursory signals rather than methodological artifacts. Given this robustness, we adopt the 2-year backward median filter for the remainder of our analysis. Influence of the acceptance threshold The acceptance threshold defines the minimum percentage of valid (cloud-free) pixels required for a scene to be included in the time series. As this threshold increases, a larger fraction of scenes is rejected due to insufficient valid data, leading to more temporal gaps in the resulting record. Consequently, higher thresholds increase the reliance on interpolation to fill missing values, potentially introducing greater uncertainty and smoothing of short-term variations. For instance, when using acceptance thresholds of 0, 1 and 5% for the summit area, the corresponding fraction of gaps is ~ 40% (largest gap = 18 days; average number of days with data per year = 221 days), ~ 46% (largest gap = 18 days; average number of days with data per year = 198 days), and ~ 55% (largest gap = 24 days; average number of days with data per year = 164 days), respectively (Figs. 8b and 9). This illustrates that even modest increases in the acceptance threshold can substantially reduce data availability and continuity, particularly for small regions of interest (ROIs). In turn, this reduction can influence the stability of the retrieved trends. For example, while a 1% threshold (Fig. 9a) still captures clear pre-eruptive increases in thermal anomalies—similar to those obtained with a 0% threshold (Fig. 8b)—a 5% threshold (Fig. 9b) begins to obscure the summit-centered pre-eruptive signals, especially those preceding the 2008 and 2019–2020 events. The sensitivity of data retention to threshold choice is especially pronounced for small ROIs: in the case considered here (16 km2, i.e., up to 16 pixels), the limited number of valid pixels amplifies the impact of thresholding on both data continuity and interpolation. By contrast, higher thresholds become more appropriate for larger ROIs, where the greater number of pixels allows spatial averaging to mitigate noise while maintaining sufficient data coverage.Fig. 9Results obtained with SSTAR for Shishaldin volcano using different acceptance thresholds. a Results for 1% acceptance threshold. b Results for 5% acceptance threshold. In both cases, we use an ROI centered at coordinates (0, 0) with a 4 km side length, linear interpolation, data acquired under both daytime and nighttime conditions, low-pass (median) filter with 730-day cutoff period, and 100 iterations for uncertainty quantification. The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) Results obtained with SSTAR for Shishaldin volcano using different acceptance thresholds. a Results for 1% acceptance threshold. b Results for 5% acceptance threshold. In both cases, we use an ROI centered at coordinates (0, 0) with a 4 km side length, linear interpolation, data acquired under both daytime and nighttime conditions, low-pass (median) filter with 730-day cutoff period, and 100 iterations for uncertainty quantification. The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) Check for potential spurious long-term trends A key concern in interpreting diffuse, low-temperature thermal anomaly time series is the possibility that long-term, non-volcanic trends remain insufficiently removed during the differencing and filtering process, thereby introducing spurious signals that could be mistaken for genuine precursory changes. To assess potential contamination effects, we compare the anomaly time series (Fig. 8b) with the temporal evolution of the median brightness temperature of both the ROI (Fig. 10a) and the Reference area (Fig. 10b). This approach allows us to evaluate whether the pre-eruptive trends identified in the anomaly record might simply reflect residual long-term variability that also appears in the Reference dataset, which is assumed not to be affected by volcanic processes.Fig. 10Example of the thermal evolution of the region of interest (ROI) and reference, as derived with SSTAR for Shishaldin volcano. a Results for the ROI. b Results for the reference area. These plots are obtained by using an ROI centered at coordinates (0, 0) with a 4 km side length, linear interpolation, data acquired under both daytime and nighttime conditions, low-pass (median) filter with 730-day cutoff period, acceptance threshold of 0%, and 100 iterations for uncertainty quantification. The red line represents the median thermal evolution over time (note that this is not the anomaly), while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) Example of the thermal evolution of the region of interest (ROI) and reference, as derived with SSTAR for Shishaldin volcano. a Results for the ROI. b Results for the reference area. These plots are obtained by using an ROI centered at coordinates (0, 0) with a 4 km side length, linear interpolation, data acquired under both daytime and nighttime conditions, low-pass (median) filter with 730-day cutoff period, acceptance threshold of 0%, and 100 iterations for uncertainty quantification. The red line represents the median thermal evolution over time (note that this is not the anomaly), while the light red shaded area denotes the associated 95% confidence interval. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025) The Reference time series (Fig. 10b) does not exhibit any coherent behavior that aligns with eruption timings, nor does it display patterns that resemble the pre-eruptive increases observed in the anomaly curves. For instance, while both the anomaly (Fig. 8b) and the ROI (Fig. 10a) show a gradual increase beginning around 2017, the Reference (Fig. 10b) displays a decreasing trend over the same interval. Similar inconsistencies are observed across the full study period, reinforcing the absence of any shared long-term structure between the Reference and the anomaly signals. Overall, the anomaly time series for the summit area (Fig. 8b) does not mimic the variability of the Reference (Fig. 10b). This lack of correspondence strongly suggests that the pre-eruptive increases recovered through SSTAR are not artifacts arising from unremoved long-term noise or regional environmental trends. Instead, the comparison enhances confidence that the detected anomalies represent genuine volcanic-related thermal variations rather than methodological byproducts. Influence of the reference area SSTAR results represent relative variations in radiant heat emissions with respect to a user-defined reference area. This means that different reference areas, each with their own thermal evolution, might yield different values of thermal anomalies at the ROI. To illustrate this effect and assess the variability of results at Shishaldin, we examined three additional configuration setups (Fig. 11) beyond the primary configuration used throughout the manuscript (Figs. 4 and corresponding results in Fig. 8b).Fig. 11Results obtained with SSTAR for Shishaldin volcano using different reference areas. a Region of interest (ROI; in blue) of 16 km2 centered at the summit of Shishaldin volcano with a reference area of 50 km internal side and 80 km external side (green). b Results obtained for configuration setup in panel a. c ROI (blue) of 16 km2 centered at the summit of Shishaldin volcano with a reference area of 40 km internal side and 80 km external side (green). d Results obtained for configuration setup in panel c. e Region of interest (ROI; in blue) of 16 km2 centered at the summit of Shishaldin volcano with a reference area formed by a square of 10 km side located at the feet of the volcano (green). f Results obtained for configuration setup in panel e. Compare these configuration setups and results with Figs. 4 and 8b. The red points in panels a, c, and e represent the target locations used for the analysis. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025). Results are for 0% acceptance threshold, linear interpolation, daytime and nighttime conditions, low-pass (median) filter with 730-day cutoff period, and 100 iterations for uncertainty quantification. The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) Results obtained with SSTAR for Shishaldin volcano using different reference areas. a Region of interest (ROI; in blue) of 16 km2 centered at the summit of Shishaldin volcano with a reference area of 50 km internal side and 80 km external side (green). b Results obtained for configuration setup in panel a. c ROI (blue) of 16 km2 centered at the summit of Shishaldin volcano with a reference area of 40 km internal side and 80 km external side (green). d Results obtained for configuration setup in panel c. e Region of interest (ROI; in blue) of 16 km2 centered at the summit of Shishaldin volcano with a reference area formed by a square of 10 km side located at the feet of the volcano (green). f Results obtained for configuration setup in panel e. Compare these configuration setups and results with Figs. 4 and 8b. The red points in panels a, c, and e represent the target locations used for the analysis. The vertical blue bands mark the onset and duration of eruptive events, highlighting periods of documented volcanic activity according to the Global Volcanism Program (Andrews et al. 2025). Results are for 0% acceptance threshold, linear interpolation, daytime and nighttime conditions, low-pass (median) filter with 730-day cutoff period, and 100 iterations for uncertainty quantification. The red line represents the median thermal anomaly over time, while the light red shaded area denotes the associated 95% confidence interval. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) When different annular-square reference regions are placed around the volcano (Fig. 11), the resulting anomaly time series show no substantial deviations from the main configuration (Figs. 4 and 8b). The overall evolution of the ROI remains consistent, and the key pre-eruptive features persist across all of these tests. More noticeable differences emerge when the ROI is compared against a 100 km2 reference area located at the base of the volcanic edifice (Fig. 11e), yet the pre-eruptive anomalies remain evident (Fig. 11f). Relative to this low-elevation reference area, the diffuse thermal anomalies display a greater amplitude, with the most recent values showing increases by roughly 4 K since 2017. In addition, the anomaly does not decrease following the unconfirmed 2008 event, but the pre-eruptive rise preceding the 2019–2020 eruption becomes more pronounced. These comparisons show that, for Shishaldin, the main pre-eruptive signatures are robust across a range of reference configurations. At the same time, they underscore that diffuse thermal anomalies always refer to the chosen reference region and must therefore be interpreted in that context. Overall, Shishaldin provides consistent, stable results even under substantial variations in reference-area geometry and location, reinforcing confidence in the suitability of SSTAR for routine diffuse thermal unrest monitoring. Spatial distribution of the thermal anomalies One of the key strengths of SSTAR, and a major advancement over the approach proposed by Girona et al. (2021), is its ability not only to track diffuse thermal anomalies at the summit or across the whole volcanic edifice, but also to generate spatially resolved thermal anomaly maps, which enable the detection and localization of potential precursory signals throughout the entire edifice. Building on this enhanced capability, we examined whether anomalies developed beyond the summit region by computing the spatiotemporal evolution of low-temperature thermal anomalies during the 15 months preceding the 2008, 2014–2016, 2019–2020, and 2023 eruptions. For each event, we selected as a temporal reference an early stage of the pre-eruptive trend identified in the summit-centered anomaly curve (Fig. 8b). Hence, this approach allows us to examine when and where thermal unrest intensified during the run-up to each eruption, providing a more comprehensive view of the spatial distribution and evolution of the anomalies prior to eruptive activity. A cross-event comparison shows that all eruptions follow a broadly consistent pattern of progressive thermal escalation, yet substantial differences arise in both the timing and the spatial distribution of their precursory signals. Specifically, the 2008 eruption (Fig. 12) was preceded by a pronounced and steadily intensifying anomalous episode on the southern flank of the volcano, which evolved systematically over the six months leading up to the eruption. Thermal anomalies increased markedly in magnitude during the final days before the event and extended to ~ 5 km from the summit, indicating a widespread thermal response across the edifice. Unexpectedly, two additional high-intensity anomalies developed as well on the flanks at considerable distances from the summit—one located approximately 9 km west and 1 km north of the crater, and another about 7 km east and 3 km north. Both sites exhibited substantial thermal increases several months prior to the eruption and maintained elevated values throughout the six-month pre-eruptive period.Fig. 12Spatiotemporal distribution of low-temperature thermal anomalies preceding the February 12, 2008 eruption of Shishaldin volcano. Spatial distribution of diffuse thermal anomalies on a November 12, 2006; b February 12, 2007; c May 12, 2007; d August 12, 2007; e November 12, 2007; and f February 11, 2008, i.e., 15 months, 12 months, 9 months, 6 months, 3 months, and 1 day, respectively, before the eruption. These results are derived from the analysis of 221 regions of interest (ROIs) distributed around the volcano center, arranged in two grids with 2 km spacing (see Fig. 4) and with each ROI covering an area of 4 × 4 km. The dataset combines observations acquired under both daytime and nighttime conditions and was processed using linear temporal interpolation, a low-pass median filter with a 730-day cutoff period, and acceptance threshold of 0%. Linear spatial interpolation across a 400-point spatial grid and 0.7% transparency were then applied to produce the final map. Colors represent the magnitude of the thermal anomaly, while negative values are excluded to enhance visualization. The temporal reference for calculating the distribution of anomalies is set to August 12, 2006, i.e., 18 months before the eruption. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) Spatiotemporal distribution of low-temperature thermal anomalies preceding the February 12, 2008 eruption of Shishaldin volcano. Spatial distribution of diffuse thermal anomalies on a November 12, 2006; b February 12, 2007; c May 12, 2007; d August 12, 2007; e November 12, 2007; and f February 11, 2008, i.e., 15 months, 12 months, 9 months, 6 months, 3 months, and 1 day, respectively, before the eruption. These results are derived from the analysis of 221 regions of interest (ROIs) distributed around the volcano center, arranged in two grids with 2 km spacing (see Fig. 4) and with each ROI covering an area of 4 × 4 km. The dataset combines observations acquired under both daytime and nighttime conditions and was processed using linear temporal interpolation, a low-pass median filter with a 730-day cutoff period, and acceptance threshold of 0%. Linear spatial interpolation across a 400-point spatial grid and 0.7% transparency were then applied to produce the final map. Colors represent the magnitude of the thermal anomaly, while negative values are excluded to enhance visualization. The temporal reference for calculating the distribution of anomalies is set to August 12, 2006, i.e., 18 months before the eruption. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) The thermal pattern preceding the 2014–2015 eruption was dominated by the emergence of a prominent anomaly on the northwestern flank of the volcano (Fig. 13). This feature extended continuously from the summit down to nearly 7 km along the edifice, suggesting a broad zone of elevated thermal background in the lead-up to the eruption. In contrast, the southern-flank anomaly—clearly expressed prior to the 2008 event—was much weaker during this period. Notably, another warmer area developed on the western flank, very close to the location of one of the distal anomalies detected before the 2008 eruption, and it showed a sustained and progressive increase in thermal emissions within the year preceding the event. Apparent anomalies at the edges of the thermal maps do not represent real thermal changes; rather, they are likely artifacts arising from spatial interpolation.Fig. 13Spatiotemporal distribution of low-temperature thermal anomalies preceding the January 30, 2014–July 3, 2015 eruption of Shishaldin volcano. Spatial distribution of diffuse thermal anomalies on a October 30, 2012; b January 30, 2013; c April 30, 2013; d July 30, 2013; e October 30, 2013; and f January 29, 2014, i.e., 15 months, 12 months, 9 months, 6 months, 3 months, and 1 day, respectively, before the onset of the eruption. These results are derived from the analysis of 221 regions of interest (ROIs) distributed around the volcano center, arranged in two grids with 2 km spacing (see Fig. 4) and with each ROI covering an area of 4 × 4 km. The dataset combines observations acquired under both daytime and nighttime conditions and was processed using linear temporal interpolation, a low-pass median filter with a 730-day cutoff period, and acceptance threshold of 0%. Linear spatial interpolation across a 400-point spatial grid and 0.7% transparency were then applied to produce the final map. Colors represent the magnitude of the thermal anomaly, while negative values are excluded to enhance visualization. The temporal reference for calculating the anomalies is set to July 30, 2012, i.e., 18 months before the onset of the eruption. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) Spatiotemporal distribution of low-temperature thermal anomalies preceding the January 30, 2014–July 3, 2015 eruption of Shishaldin volcano. Spatial distribution of diffuse thermal anomalies on a October 30, 2012; b January 30, 2013; c April 30, 2013; d July 30, 2013; e October 30, 2013; and f January 29, 2014, i.e., 15 months, 12 months, 9 months, 6 months, 3 months, and 1 day, respectively, before the onset of the eruption. These results are derived from the analysis of 221 regions of interest (ROIs) distributed around the volcano center, arranged in two grids with 2 km spacing (see Fig. 4) and with each ROI covering an area of 4 × 4 km. The dataset combines observations acquired under both daytime and nighttime conditions and was processed using linear temporal interpolation, a low-pass median filter with a 730-day cutoff period, and acceptance threshold of 0%. Linear spatial interpolation across a 400-point spatial grid and 0.7% transparency were then applied to produce the final map. Colors represent the magnitude of the thermal anomaly, while negative values are excluded to enhance visualization. The temporal reference for calculating the anomalies is set to July 30, 2012, i.e., 18 months before the onset of the eruption. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) The 2019–2020 eruption exhibits the least pronounced pre-eruptive anomaly at the summit, consistent with the behavior observed in the time series (Fig. 8). Despite this weak summit expression, the pre-eruptive phase is characterized by the emergence of several well-defined anomalies on the northern flank of the volcano, located at considerable distances from the crater (Fig. 14). In particular, a strong anomaly developed approximately 6 km north of the summit, accompanied by another broad diffuse anomaly area situated about 3 km east and 8 km north of the summit. A weaker but still notable anomaly also appeared around 8 km west and 8 km north. These distal thermal signals indicate that the 2019–2020 unrest followed a distributed and spatially complex thermal footprint, with the most significant anomalies occurring far from the summit region.Fig. 14Spatiotemporal distribution of low-temperature thermal anomalies preceding the July 23, 2019–May 4, 2020 eruption of Shishaldin volcano. Spatial distribution of diffuse thermal anomalies on a April 23, 2018; b July 23, 2018; c October 23, 2018; d January 23, 2019; e April 23, 2019; and f July 22, 2019, i.e., 15 months, 12 months, 9 months, 6 months, 3 months, and 1 day, respectively, before the onset of the eruption. These results are derived from the analysis of 221 regions of interest (ROIs) distributed around the volcano center, arranged in two grids with 2 km spacing (see Fig. 4) and with each ROI covering an area of 4 × 4 km. The dataset combines observations acquired under both daytime and nighttime conditions and was processed using linear temporal interpolation, a low-pass median filter with a 730-day cutoff period, and acceptance threshold of 0%. Linear spatial interpolation across a 400-point spatial grid and 0.7% transparency were then applied to produce the final map. Colors represent the magnitude of the thermal anomaly, while negative values are excluded to enhance visualization. The temporal reference for calculating the anomalies is set to January 23, 2018, i.e., 18 months before the onset of the eruption. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) Spatiotemporal distribution of low-temperature thermal anomalies preceding the July 23, 2019–May 4, 2020 eruption of Shishaldin volcano. Spatial distribution of diffuse thermal anomalies on a April 23, 2018; b July 23, 2018; c October 23, 2018; d January 23, 2019; e April 23, 2019; and f July 22, 2019, i.e., 15 months, 12 months, 9 months, 6 months, 3 months, and 1 day, respectively, before the onset of the eruption. These results are derived from the analysis of 221 regions of interest (ROIs) distributed around the volcano center, arranged in two grids with 2 km spacing (see Fig. 4) and with each ROI covering an area of 4 × 4 km. The dataset combines observations acquired under both daytime and nighttime conditions and was processed using linear temporal interpolation, a low-pass median filter with a 730-day cutoff period, and acceptance threshold of 0%. Linear spatial interpolation across a 400-point spatial grid and 0.7% transparency were then applied to produce the final map. Colors represent the magnitude of the thermal anomaly, while negative values are excluded to enhance visualization. The temporal reference for calculating the anomalies is set to January 23, 2018, i.e., 18 months before the onset of the eruption. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) The thermal pattern preceding the 2023 eruption was considerably simpler than those associated with previous eruptive episodes (Fig. 15). Most of the anomaly was focused on the summit crater and along the southern flank, a distribution that closely mirrors the configuration observed prior to the 2008 eruption. This localized phase of thermal unrest intensified during the final three months before the onset of the eruption and followed a well-defined, roughly west–east alignment extending from approximately 4 km west of the summit to about 1 km east, with a slight southward offset relative to the crater. Beyond this narrow alignment, no significant or coherent thermal anomalies were detected, indicating that the precursory activity in 2023 was restricted to a relatively limited portion of the edifice.Fig. 15Spatiotemporal distribution of low-temperature thermal anomalies preceding the July 11–November 3, 2023 eruption of Shishaldin volcano. Spatial distribution of diffuse thermal anomalies on a April 11, 2022; b July 11, 2022; c October 11, 2022; d January 11, 2023; e April 11, 2023; and f July 10, 2023, i.e., 15 months, 12 months, 9 months, 6 months, 3 months, and 1 day, respectively, before the onset of the eruption. These results are derived from the analysis of 221 regions of interest (ROIs) distributed around the volcano center, arranged in two grids with 2 km spacing (see Fig. 4) and with each ROI covering an area of 4 × 4 km. The dataset combines observations acquired under both daytime and nighttime conditions and was processed using linear temporal interpolation, a low-pass median filter with a 730-day cutoff period, and acceptance threshold of 0%. Linear spatial interpolation across a 400-point spatial grid and 0.7% transparency were then applied to produce the final map. Colors represent the magnitude of the thermal anomaly, while negative values are excluded to enhance visualization. The temporal reference for calculating the anomalies is set to March 30, 2022, i.e., ~ 12.5 months before the onset of the eruption, when the pre-eruptive increasing trend at the summit begins. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) Spatiotemporal distribution of low-temperature thermal anomalies preceding the July 11–November 3, 2023 eruption of Shishaldin volcano. Spatial distribution of diffuse thermal anomalies on a April 11, 2022; b July 11, 2022; c October 11, 2022; d January 11, 2023; e April 11, 2023; and f July 10, 2023, i.e., 15 months, 12 months, 9 months, 6 months, 3 months, and 1 day, respectively, before the onset of the eruption. These results are derived from the analysis of 221 regions of interest (ROIs) distributed around the volcano center, arranged in two grids with 2 km spacing (see Fig. 4) and with each ROI covering an area of 4 × 4 km. The dataset combines observations acquired under both daytime and nighttime conditions and was processed using linear temporal interpolation, a low-pass median filter with a 730-day cutoff period, and acceptance threshold of 0%. Linear spatial interpolation across a 400-point spatial grid and 0.7% transparency were then applied to produce the final map. Colors represent the magnitude of the thermal anomaly, while negative values are excluded to enhance visualization. The temporal reference for calculating the anomalies is set to March 30, 2022, i.e., ~ 12.5 months before the onset of the eruption, when the pre-eruptive increasing trend at the summit begins. Satellite background imagery for maps was obtained from the Esri World Imagery basemap, accessed through MATLAB R2021b (MathWorks, Natick, USA) Discussion In this section, we first situate SSTAR and diffuse thermal anomalies within the context of previous work, which has largely focused on hotspot detection. We then discuss the role of snow and ice cover, followed by a discussion of the Shishaldin results and their broader implications. We next outline key challenges and recommendations for the effective application of SSTAR. Volcanic hotspots vs diffuse thermal anomalies Multiple methods and platforms have been developed over the past two decades to detect and analyze thermal anomalies using satellite-based infrared data (Wright et al. 2004; Ganci et al. 2011; Lacava et al. 2011; Coppola et al. 2016; Gouhier et al. 2016; Lombardo 2016; Pergola et al. 2016; Gordeev et al. 2016; Genzano et al. 2020; Layana et al. 2020; Mazzeo et al. 2021; Torrisi et al. 2022; Corradino et al. 2023; Amato et al. 2023; Cariello et al. 2024; Saunders-Shultz et al. 2024). Most of these approaches rely on calculating differences in radiant heat emission, either between a volcanic target and its surroundings or between successive observations, to identify deviations from background thermal conditions; this enables the detection of hotspots and the construction of time series of volcanic radiant power or thermal excess for monitoring eruptive activity. These techniques have proven effective for identifying strong, localized thermal signals associated with high-temperature features such as lava flows, domes, fumarolic activity, or hot vents. In contrast, SSTAR is specifically designed to enable the systematic detection and analysis of low-amplitude, spatially coherent thermal anomalies over broad areas and long timescales (Girona et al. 2021; Girona and Brenot 2026). Rather than targeting individual hotspots, SSTAR quantifies relative radiant heat variations over time by comparing the median brightness temperature of a region of interest (ROI), typically spanning several square kilometers, with that of a carefully selected reference area. This strategy enables the detection of very subtle, spatially distributed signals that reflect the surface expression of subsurface unrest processes. These subsurface processes may include shallow phase changes and hydrothermal circulation (García-Martínez et al. 2025), while their surface manifestation might involve actual ground warming or indirect effects such as emissivity changes caused by mass movements or stressed vegetation (Cawse-Nicholson et al. 2018; Bogue et al. 2019, 2023). Accessibility and reproducibility represent an additional point of distinction between SSTAR and previous algorithms. Existing algorithms have been implemented within specific software environments or described only at a general methodological level, which can hinder systematic, community-wide application. In contrast, SSTAR is, to the best of our knowledge, the first open, freely available, and user-friendly framework dedicated to the detailed analysis of satellite-derived thermal anomalies in volcanic and hydrothermal environments. Its design enables straightforward use by both researchers and monitoring agencies, helping to promote the standardization of thermal time-series analysis across a wide range of volcanic and geothermal settings. The algorithm also incorporates all the processing steps, such as cloud filtering, temporal interpolation, and denoising, in a transparent and well-documented manner, thereby enhancing the interpretability of long-term thermal trends while ensuring clarity in data assessment and full user control over the analysis. The role of snow and ice cover on the detection of diffuse thermal anomalies Snow and ice cover are ubiquitous at many active volcanoes worldwide, particularly in high-latitude and high-altitude regions such as Alaska, Iceland, and the Andes. Variations in snow and ice thickness and extent can significantly affect both surface emissivity and physical temperature, and therefore influence satellite-derived brightness temperature in ways that are not directly related to subsurface volcanic or hydrothermal processes, potentially complicating the interpretation of diffuse anomalies. Within the SSTAR framework, these influences are mitigated through the combined use of relative analysis and temporal filtering. SSTAR evaluates thermal variations within a region of interest (ROI) relative to a reference area. When both the ROI and the reference experience similar atmospheric and surface conditions, including comparable snow or ice cover, these effects are largely canceled out, as the method focuses on relative differences rather than absolute values. This approach effectively reduces the impact of spatially coherent environmental signals; however, in many practical applications, the reference area cannot be located immediately adjacent to the ROI, and differences in elevation or local climate may result in distinct snow and ice conditions between the two areas. This decoupling can introduce variability in the relative signal, expressed as short-term fluctuations and seasonal trends in the retrieved anomaly time series. In SSTAR, such effects are mitigated through the application of temporal low-pass filtering (e.g., a 2-year window), which attenuates seasonal and higher-frequency variability while preserving longer-term, persistent trends associated with subsurface unrest. Overall, although snow and ice cover can influence thermal observations, their impact on SSTAR-derived diffuse thermal anomalies is substantially minimized through the combined use of relative comparison and temporal filtering. This enables the robust detection of subtle, long-term thermal signals even in environments characterized by significant snow and ice variability. Link between diffuse thermal anomalies and unrest at Shishaldin volcano, Alaska In this subsection, we summarize the main findings for Shishaldin volcano and outline feasible interpretations based on our July 4, 2002–December 31, 2024 SSTAR analysis. It is important to emphasize, however, that this manuscript is a technical report focused on presenting the SSTAR framework and illustrating its application through a representative case study. Accordingly, our analysis of Shishaldin is not intended to be an exhaustive interpretation of its diffuse thermal emissions or pre-eruptive processes. For a more detailed examination of how satellite-derived diffuse thermal anomalies relate to independent geophysical and geological observations, as well as to complementary remote sensing datasets and the underlying physical processes driving unrest at Shishaldin, we refer readers to Angarita et al. (2026). Our findings and interpretations are:(i)The ~ 16 km2 region surrounding the summit, i.e., the upper sector of the volcanic edifice, consistently exhibits pre-eruptive diffuse thermal anomalies, displaying a sequence of thermal phases that closely mirrors the independently documented eruptive history of Shishaldin (Fig. 8b). Each eruption was preceded by a significant increase in the low-temperature anomaly (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\updelta \mathrm{T} }$$\end{document}δT¯), followed by a gradual decline as activity wanes, while quiescent periods show decreasing or low values, typically below ~ 1 K. A particularly compelling case is the 2008 episode; although listed as “unconfirmed” in the Global Volcanism Program database (Andrews et al. 2025), SSTAR reveals a pronounced increase in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\updelta \mathrm{T} }$$\end{document}δT¯ within the ~ 16 km2 ROI around the summit. The magnitude of this anomaly (~ 1.3 K) is comparable to those observed before confirmed eruptions, suggesting that the volcano experienced genuine thermal unrest at that time. The strong thermal sensitivity of the upper edifice of Shishaldin to pre-eruptive processes, even ahead of small eruptions, makes it an especially valuable sector for near–real-time monitoring of diffuse thermal anomalies and for supporting alert-level assessments and eruption forecasting efforts.(ii)Several well-defined flank locations, some more than 5 km from the summit, also exhibit marked diffuse thermal anomalies prior to specific events, suggesting that these areas may also reflect, directly or indirectly, the preparatory phase of eruptions. Before the 2008 episode, for example, SSTAR detects a ~ 3–4 km anomalous alignment roughly 9 km west of the summit (Fig. 12), overlapping with small cinder cones near the base of the volcano and close to the ice limit. Likewise, the 2014–2015 event was preceded by an anomalous thermal pattern on the northwestern flank (Fig. 13), coinciding with a dense cluster of cinder cones; indeed, more than 50 cinder cones have been identified on Shishaldin flanks, most of them in this sector, where phreatomagmatic explosions have also occurred in the past (Beget et al. 2003). Because cinder cones and related volcanotectonic structures often correspond to zones of enhanced crustal permeability (Gudmundsson 2022), diffuse thermal anomalies detected in these areas may reflect the preferential circulation of relatively warm volatiles or fluids along structurally controlled pathways, potentially driven by shallow magma migrations (Lundgren et al. 2020; Zhan et al. 2022). This interpretation is consistent with the observation that the temporal evolution of flank anomalies often mirrors that of the upper edifice. However, the strong anomalies detected ~ 6–7 km north of the summit prior to the 2019–2020 eruption (Fig. 14) are more difficult to interpret without complementary datasets. Although their origin is uncertain, their temporal correspondence with volcanic unrest suggests they might still be indirectly linked to pre-eruptive processes; for example, mass movements triggered by partial melting of snow or glacier ice (Howcutt et al. 2023; Martin et al. 2025; Unnsteinsson et al. 2025).(iii)Pre-eruptive diffuse thermal anomalies may represent genuine temperature increases (García-Martínez et al. 2025), but they might also reflect, to some extent, emissivity changes resulting from surface modifications driven by unrest. This arises from the fact that SSTAR uses brightness temperature as a proxy for thermal infrared anomalies, minimizing over-processing to preserve subtle trends (Girona et al. 2021). Unrest-driven surface emissivity changes may reflect vegetation stress (Cawse-Nicholson et al. 2018; Bogue et al. 2019, 2023), which might be relevant on the lower flanks of Shishaldin; and cryospheric changes, such as reduced snow or ice cover, triggered by partial melting (Howcutt et al. 2023; Martin et al. 2025; Unnsteinsson et al. 2025). Future studies should aim to further constrain the processes driving diffuse thermal anomalies and to clarify how potential unrest-related changes in vegetation, snow, and ice might contribute to the observed pre-eruptive thermal signals.(iv)The intensity of pre-eruptive diffuse thermal anomalies does not appear to correlate with eruption size. For example, the 2019–2020 VEI 3 eruption was preceded by a relatively modest anomaly in the upper edifice (< 1 K), whereas the smaller (and officially “unconfirmed”) 2008 event showed a markedly stronger thermal signal (Fig. 8b). This contrast suggests that the amplitude of diffuse thermal unrest is governed primarily by shallow adjustments in subsurface heat and fluid circulation rather than by the eventual magnitude of the eruption. Accordingly, while diffuse thermal anomalies provide useful information on the timing and evolution of unrest, their magnitude should not be interpreted as an indicator of a forthcoming eruption size. In any case, this potential link between pre-eruptive, diffuse thermal anomalies and eruption size should be revisited as more eruptions become available for analysis.(v)Interestingly, for the two longest eruptions (the 2014–2015 and 2019–2020 events), the diffuse thermal trends retrieved for the ~ 16 km2 region surrounding the summit begin to decline almost immediately at eruption onset, rather than near the end of activity (Fig. 8b). A plausible explanation is that the fluids responsible for generating diffuse thermal signals (García-Martínez et al. 2025) cease circulating in the shallow subsurface of the flanks once an eruption begins; instead, they are rapidly redirected and focused through newly opened, highly permeable eruptive vents. This interpretation is consistent with the SSTAR results, as the method is intended to detect spatially distributed thermal anomalies but is not optimized to capture localized, high-temperature sources such as small eruptive vents. (i)The ~ 16 km2 region surrounding the summit, i.e., the upper sector of the volcanic edifice, consistently exhibits pre-eruptive diffuse thermal anomalies, displaying a sequence of thermal phases that closely mirrors the independently documented eruptive history of Shishaldin (Fig. 8b). Each eruption was preceded by a significant increase in the low-temperature anomaly (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\updelta \mathrm{T} }$$\end{document}δT¯), followed by a gradual decline as activity wanes, while quiescent periods show decreasing or low values, typically below ~ 1 K. A particularly compelling case is the 2008 episode; although listed as “unconfirmed” in the Global Volcanism Program database (Andrews et al. 2025), SSTAR reveals a pronounced increase in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\updelta \mathrm{T} }$$\end{document}δT¯ within the ~ 16 km2 ROI around the summit. The magnitude of this anomaly (~ 1.3 K) is comparable to those observed before confirmed eruptions, suggesting that the volcano experienced genuine thermal unrest at that time. The strong thermal sensitivity of the upper edifice of Shishaldin to pre-eruptive processes, even ahead of small eruptions, makes it an especially valuable sector for near–real-time monitoring of diffuse thermal anomalies and for supporting alert-level assessments and eruption forecasting efforts. The ~ 16 km2 region surrounding the summit, i.e., the upper sector of the volcanic edifice, consistently exhibits pre-eruptive diffuse thermal anomalies, displaying a sequence of thermal phases that closely mirrors the independently documented eruptive history of Shishaldin (Fig. 8b). Each eruption was preceded by a significant increase in the low-temperature anomaly (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\updelta \mathrm{T} }$$\end{document}δT¯), followed by a gradual decline as activity wanes, while quiescent periods show decreasing or low values, typically below ~ 1 K. A particularly compelling case is the 2008 episode; although listed as “unconfirmed” in the Global Volcanism Program database (Andrews et al. 2025), SSTAR reveals a pronounced increase in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\updelta \mathrm{T} }$$\end{document}δT¯ within the ~ 16 km2 ROI around the summit. The magnitude of this anomaly (~ 1.3 K) is comparable to those observed before confirmed eruptions, suggesting that the volcano experienced genuine thermal unrest at that time. The strong thermal sensitivity of the upper edifice of Shishaldin to pre-eruptive processes, even ahead of small eruptions, makes it an especially valuable sector for near–real-time monitoring of diffuse thermal anomalies and for supporting alert-level assessments and eruption forecasting efforts. (ii)Several well-defined flank locations, some more than 5 km from the summit, also exhibit marked diffuse thermal anomalies prior to specific events, suggesting that these areas may also reflect, directly or indirectly, the preparatory phase of eruptions. Before the 2008 episode, for example, SSTAR detects a ~ 3–4 km anomalous alignment roughly 9 km west of the summit (Fig. 12), overlapping with small cinder cones near the base of the volcano and close to the ice limit. Likewise, the 2014–2015 event was preceded by an anomalous thermal pattern on the northwestern flank (Fig. 13), coinciding with a dense cluster of cinder cones; indeed, more than 50 cinder cones have been identified on Shishaldin flanks, most of them in this sector, where phreatomagmatic explosions have also occurred in the past (Beget et al. 2003). Because cinder cones and related volcanotectonic structures often correspond to zones of enhanced crustal permeability (Gudmundsson 2022), diffuse thermal anomalies detected in these areas may reflect the preferential circulation of relatively warm volatiles or fluids along structurally controlled pathways, potentially driven by shallow magma migrations (Lundgren et al. 2020; Zhan et al. 2022). This interpretation is consistent with the observation that the temporal evolution of flank anomalies often mirrors that of the upper edifice. However, the strong anomalies detected ~ 6–7 km north of the summit prior to the 2019–2020 eruption (Fig. 14) are more difficult to interpret without complementary datasets. Although their origin is uncertain, their temporal correspondence with volcanic unrest suggests they might still be indirectly linked to pre-eruptive processes; for example, mass movements triggered by partial melting of snow or glacier ice (Howcutt et al. 2023; Martin et al. 2025; Unnsteinsson et al. 2025). Several well-defined flank locations, some more than 5 km from the summit, also exhibit marked diffuse thermal anomalies prior to specific events, suggesting that these areas may also reflect, directly or indirectly, the preparatory phase of eruptions. Before the 2008 episode, for example, SSTAR detects a ~ 3–4 km anomalous alignment roughly 9 km west of the summit (Fig. 12), overlapping with small cinder cones near the base of the volcano and close to the ice limit. Likewise, the 2014–2015 event was preceded by an anomalous thermal pattern on the northwestern flank (Fig. 13), coinciding with a dense cluster of cinder cones; indeed, more than 50 cinder cones have been identified on Shishaldin flanks, most of them in this sector, where phreatomagmatic explosions have also occurred in the past (Beget et al. 2003). Because cinder cones and related volcanotectonic structures often correspond to zones of enhanced crustal permeability (Gudmundsson 2022), diffuse thermal anomalies detected in these areas may reflect the preferential circulation of relatively warm volatiles or fluids along structurally controlled pathways, potentially driven by shallow magma migrations (Lundgren et al. 2020; Zhan et al. 2022). This interpretation is consistent with the observation that the temporal evolution of flank anomalies often mirrors that of the upper edifice. However, the strong anomalies detected ~ 6–7 km north of the summit prior to the 2019–2020 eruption (Fig. 14) are more difficult to interpret without complementary datasets. Although their origin is uncertain, their temporal correspondence with volcanic unrest suggests they might still be indirectly linked to pre-eruptive processes; for example, mass movements triggered by partial melting of snow or glacier ice (Howcutt et al. 2023; Martin et al. 2025; Unnsteinsson et al. 2025). (iii)Pre-eruptive diffuse thermal anomalies may represent genuine temperature increases (García-Martínez et al. 2025), but they might also reflect, to some extent, emissivity changes resulting from surface modifications driven by unrest. This arises from the fact that SSTAR uses brightness temperature as a proxy for thermal infrared anomalies, minimizing over-processing to preserve subtle trends (Girona et al. 2021). Unrest-driven surface emissivity changes may reflect vegetation stress (Cawse-Nicholson et al. 2018; Bogue et al. 2019, 2023), which might be relevant on the lower flanks of Shishaldin; and cryospheric changes, such as reduced snow or ice cover, triggered by partial melting (Howcutt et al. 2023; Martin et al. 2025; Unnsteinsson et al. 2025). Future studies should aim to further constrain the processes driving diffuse thermal anomalies and to clarify how potential unrest-related changes in vegetation, snow, and ice might contribute to the observed pre-eruptive thermal signals. Pre-eruptive diffuse thermal anomalies may represent genuine temperature increases (García-Martínez et al. 2025), but they might also reflect, to some extent, emissivity changes resulting from surface modifications driven by unrest. This arises from the fact that SSTAR uses brightness temperature as a proxy for thermal infrared anomalies, minimizing over-processing to preserve subtle trends (Girona et al. 2021). Unrest-driven surface emissivity changes may reflect vegetation stress (Cawse-Nicholson et al. 2018; Bogue et al. 2019, 2023), which might be relevant on the lower flanks of Shishaldin; and cryospheric changes, such as reduced snow or ice cover, triggered by partial melting (Howcutt et al. 2023; Martin et al. 2025; Unnsteinsson et al. 2025). Future studies should aim to further constrain the processes driving diffuse thermal anomalies and to clarify how potential unrest-related changes in vegetation, snow, and ice might contribute to the observed pre-eruptive thermal signals. (iv)The intensity of pre-eruptive diffuse thermal anomalies does not appear to correlate with eruption size. For example, the 2019–2020 VEI 3 eruption was preceded by a relatively modest anomaly in the upper edifice (< 1 K), whereas the smaller (and officially “unconfirmed”) 2008 event showed a markedly stronger thermal signal (Fig. 8b). This contrast suggests that the amplitude of diffuse thermal unrest is governed primarily by shallow adjustments in subsurface heat and fluid circulation rather than by the eventual magnitude of the eruption. Accordingly, while diffuse thermal anomalies provide useful information on the timing and evolution of unrest, their magnitude should not be interpreted as an indicator of a forthcoming eruption size. In any case, this potential link between pre-eruptive, diffuse thermal anomalies and eruption size should be revisited as more eruptions become available for analysis. The intensity of pre-eruptive diffuse thermal anomalies does not appear to correlate with eruption size. For example, the 2019–2020 VEI 3 eruption was preceded by a relatively modest anomaly in the upper edifice (< 1 K), whereas the smaller (and officially “unconfirmed”) 2008 event showed a markedly stronger thermal signal (Fig. 8b). This contrast suggests that the amplitude of diffuse thermal unrest is governed primarily by shallow adjustments in subsurface heat and fluid circulation rather than by the eventual magnitude of the eruption. Accordingly, while diffuse thermal anomalies provide useful information on the timing and evolution of unrest, their magnitude should not be interpreted as an indicator of a forthcoming eruption size. In any case, this potential link between pre-eruptive, diffuse thermal anomalies and eruption size should be revisited as more eruptions become available for analysis. (v)Interestingly, for the two longest eruptions (the 2014–2015 and 2019–2020 events), the diffuse thermal trends retrieved for the ~ 16 km2 region surrounding the summit begin to decline almost immediately at eruption onset, rather than near the end of activity (Fig. 8b). A plausible explanation is that the fluids responsible for generating diffuse thermal signals (García-Martínez et al. 2025) cease circulating in the shallow subsurface of the flanks once an eruption begins; instead, they are rapidly redirected and focused through newly opened, highly permeable eruptive vents. This interpretation is consistent with the SSTAR results, as the method is intended to detect spatially distributed thermal anomalies but is not optimized to capture localized, high-temperature sources such as small eruptive vents. Interestingly, for the two longest eruptions (the 2014–2015 and 2019–2020 events), the diffuse thermal trends retrieved for the ~ 16 km2 region surrounding the summit begin to decline almost immediately at eruption onset, rather than near the end of activity (Fig. 8b). A plausible explanation is that the fluids responsible for generating diffuse thermal signals (García-Martínez et al. 2025) cease circulating in the shallow subsurface of the flanks once an eruption begins; instead, they are rapidly redirected and focused through newly opened, highly permeable eruptive vents. This interpretation is consistent with the SSTAR results, as the method is intended to detect spatially distributed thermal anomalies but is not optimized to capture localized, high-temperature sources such as small eruptive vents. Overall, the Shishaldin case study demonstrates that SSTAR captures meaningful patterns of diffuse thermal unrest across multiple eruptive episodes and spatial domains. These results reinforce the value of long-term, low-temperature thermal anomalies as indicators of pre-eruptive activity and show that integrating both summit and flank signals provides a more complete picture of eruption preparatory processes. Challenges and good practices SSTAR can reveal long-term, low-amplitude thermal variations that would otherwise remain undetected, allowing subtle precursory trends to be identified and providing new insights into the dynamics of volcanic and geothermal systems. At the same time, interpreting these outputs requires careful attention to the physical environment and the observational limitations of satellite-based thermal data. The analysis of diffuse thermal anomalies introduces several methodological and practical challenges that must be addressed to ensure robust and reliable results. The following subsections summarize key considerations and recommended practices that emerged from the development, implementation, and testing of the SSTAR framework. Data volume and computational demands The analyses of diffuse thermal anomalies often rely on two decades or more of MODIS Aqua and Terra observations. Because each volcano requires retrieving, storing, and processing a very large number of scenes, the total data volume can reach several terabytes. Managing this volume efficiently is therefore a key requirement for conducting large-scale thermal studies with SSTAR. Efficient organization begins with how the data are stored. Keeping raw satellite imagery separate from intermediate files and from final outputs—such as thermal anomaly composites, anomaly time series, and figures—helps maintain performance and reduces the chance of overwriting files. This separation also supports reproducibility and version control, both of which are essential for long-term monitoring where preprocessing steps, algorithms, or calibration datasets may change. SSTAR follows this principle by automatically generating a structured directory layout that keeps all intermediate and final files clearly organized, allowing the analysis to remain transparent and easy to verify. Storage choices further influence performance. Raw MODIS scenes, which occupy most of the disk space, are best stored on high-capacity local or external drives. Processed outputs, by contrast, are smaller but more frequently accessed, and therefore benefit from faster solid-state storage. When analyses expand to multiple volcanoes or broader regional grids, using multi-core workstations, local clusters, or high-performance computing becomes advantageous because independent spatial tiles or periods of time can be processed concurrently. In all cases, maintaining comprehensive metadata—including data sources, preprocessing steps, quality-control thresholds, and file versions—is essential for ensuring traceability and long-term continuity. SSTAR generates these metadata files automatically during processing, helping maintain consistency across analyses and supporting the continuity required for operational thermal monitoring. Small regions of interest (ROIs) and/or reference area If the region of interest (ROI) or the reference area is relatively small, two practical limitations may arise. The first is an increased sensitivity to data gaps caused by cloud-contaminated pixels. When many pixels within the ROI or the reference area are affected by cloud cover and removed during quality filtering, the proportion of missing observations increases, making the resulting time series more dependent on the interpolation scheme used to fill these temporal gaps. Although interpolation is necessary to maintain continuity, relying too heavily on it can distort subtle thermal trends when the underlying data are sparse. A second limitation arises from the influence of localized high-temperature anomalies. Small ROIs or reference areas can be disproportionately affected by features such as fumaroles, small lava flows, or other discrete heat sources, even though SSTAR is not designed to capture hotspots. Although SSTAR relies on the median, a statistic that reduces the influence of outliers by reflecting the central tendency of pixel values, one or two anomalously hot pixels can bias the time series when the ROI or reference area are too small. In such cases, short-lived or spatially confined hotspots may dominate the signal and obscure the low-amplitude variations that characterize genuine diffuse thermal unrest. Addressing these issues benefits from enlarging the ROI and/or the reference areas when feasible, as larger spatial domains reduce sensitivity to both data gaps and isolated hot pixels. Evaluating multiple ROIs of different sizes, shapes, or positions can further help distinguish regional diffuse signals from localized anomalies, and determine an appropriate ROI and reference size. Despite these challenges, our results show that diffuse thermal anomalies can still be robustly detected in relatively small ROIs when data gaps remain within reasonable limits. At Shishaldin, for instance, data gaps reach up to ~ 40%, yet preeruptive signals were consistently recovered using an ROI of ~ 16 km2. This shows that SSTAR’s median-based workflow, including its cloud filtering, interpolation, and denoising steps, can effectively extract long-term thermal trends even with modest ROI sizes. Nonetheless, analyses based on small ROIs should always be accompanied by spatial diagnostics to verify that the observed variations are coherent and not driven by isolated pixels, transient high-temperature features, or spatial interpolation artifacts. Extended temporal record and data sampling Reliable analyses with SSTAR require an extended temporal record. In practice, we recommend using the full MODIS archive (over 20 years), as it provides a stable baseline for identifying deviations from background behavior and reduces the likelihood that short-term variability is misinterpreted as genuine unrest. High temporal sampling is also especially important for the detection of diffuse thermal anomalies. Unlike hotspot detection, which relies on strong spatial, temperature contrasts that can be captured even with sporadic observations, diffuse anomalies manifest as subtle, low-amplitude changes that develop gradually over months to years. Detecting and characterizing these slow trends requires time series with frequent, consistent observations so that they can be distinguished from noise, seasonality, and data gaps. This requirement also underscores a limitation of high-spatial-resolution satellite missions. Sensors such as ASTER provide excellent spatial detail (< 100 m), but their much lower temporal sampling (~ 2 weeks between acquisitions) is insufficient for constructing the dense, continuous time series needed for the analysis of diffuse thermal anomalies. Consequently, low-temporal-resolution satellites cannot support the type of statistical evaluation used by SSTAR, which relies on thousands of scenes to extract robust long-term thermal patterns. Looking ahead, the ideal scenario for diffuse thermal anomaly studies would combine both dense temporal sampling and sub-kilometer spatial resolution. Future satellite constellations capable of acquiring multiple thermal infrared overpasses per day at spatial resolutions finer than 1 km2 would offer a major advance. Such systems would allow the community to resolve diffuse thermal anomalies with far greater precision, reduce the reliance on interpolation, and improve our ability to detect the earliest thermal signatures of volcanic unrest. We envision that next-generation space-based thermal monitoring networks will play a central role in establishing more accurate and systematic assessments of diffuse thermal anomalies at active volcanoes. Filtering strategies A key consideration when retrieving low-temperature anomaly time series is the choice between non-causal (iterative wavelet + backward-looking median) and causal (backward-looking median) smoothing. The non-causal denoising approach is a powerful denoising technique, as wavelet-based filtering captures variability across multiple temporal scales and is highly effective at removing high-frequency noise, enhancing signal clarity (Girona et al. 2021). However, because the wavelet component uses both past and future observations to compute each value, post-eruptive signals can propagate backward in time and distort the apparent pre-eruptive signal. This sensitivity to post-eruptive products may compromise the interpretation of precursory trends if not handled carefully. In contrast, the causal denoising approach relies only on current and past observations, preserving long-term trends and producing a stable, lag-free baseline. By avoiding the use of future data, it prevents contamination from eruptive products or abrupt thermal signals that occur later in the record, ensuring a temporally consistent reconstruction of pre-eruptive signals. Its limitation, however, is that it is less effective at suppressing short-term fluctuations, which may hinder the detection of very subtle variations. Because each method has distinct strengths and limitations, they should be used in a complementary manner in diffuse thermal anomaly analyses, and neither should be applied as a default denoising approach without evaluating its impact on the specific time series under study. The complementarity of both denoising techniques is central to SSTAR’s design. Overall, we recommend that users investigating very subtle or short-lived precursory trends make use of both filtering approaches. First, users may apply the non-causal filter (iterative wavelet + backward-looking median) to obtain a clean, denoised time series, and then use the causal filter (backward-looking median) to verify temporal consistency, confirm whether an apparent signal is genuinely pre-eruptive, and estimate any temporal correction that may need to be applied to the non-causal result. For pronounced trends that clearly exceed background variability, such as those observed at Shishaldin, causal smoothing alone is often sufficient for the analysis of diffuse thermal anomalies. Conversely, in cases where trends are weak, short-lived, and embedded within high levels of noise, neither approach may reliably reveal pre-eruptive anomalies with sufficient statistical confidence. Assessing spurious long-term artifacts A critical step in evaluating the reliability of the thermal anomaly time series is assessing the presence of spurious long-term artifacts, which may arise from residual atmospheric effects or unremoved regional temperature trends. Such artifacts can mimic genuine thermal signals, potentially leading to misinterpretation of the temporal evolution of unrest or other surface processes. To identify the possible presence of these artifacts in the anomaly time series and quantify these effects, users should compare the long-term behavior of the anomaly series (obtained from the difference between the region of interest and the reference area) with that of the reference area alone. If both series display correlated long-term variations, this suggests that the retrieved signal may be dominated by regional, long-term climatic effects that have not been fully removed. Conversely, if no significant correlation exists between the two, the observed trends in the anomaly series likely reflect genuine, localized thermal variations rather than spurious background influences. This comparison provides a simple yet effective quality control step for distinguishing between true thermal unrest and artificial trends induced by broader environmental factors. Reference area The reference area is a critical component of SSTAR, and an inappropriate choice can obscure genuine volcanic signals or introduce artifacts into the anomaly time series. Its primary purpose is to suppress broad-scale, non-volcanic radiant heat fluctuations at the ROI, such as those driven by climatic, regional, or seasonal variability, so that subtle changes associated with volcanic or hydrothermal processes can be reliably isolated. This strategy assumes that the ROI and the reference area experience comparable non-volcanic fluctuations, and users should make every effort to identify a reference region that satisfies this assumption as closely as possible. In practice, finding such a reference area can at times be challenging. As a general rule, the most effective reference area is one located well beyond any hydrothermal or magmatic influence while still experiencing similar non-volcanic, radiant heat variations as the ROI. Building on this principle, we recommend using the annular-square configuration implemented in SSTAR (Fig. 4) as the default starting point, as it offers a systematic and easily reproducible way to define a distant reference area that typically yields stable results (this is analog to the configuration used in Girona et al. 2021). If this kind of configuration produces spurious long-term artifacts (see Section 5.4.5), users should explore alternative reference locations, ideally evaluating several distant candidates to identify the most stable baseline. For instance, in regions with pronounced topographic gradients, differences in elevation or slope orientation between the ROI and the reference area may lead to distinct radiative responses to external climatic perturbations. In such cases, a more suitable reference may be one that matches the ROI in elevation, aspect, and surface characteristics while still lying outside hydrothermal or magmatic influence, although finding such a reference is not always feasible. Another option is to place the reference area at the base of the volcanic edifice, provided independent information supports that this region is not expected to be affected by any kind of magmatic or hydrothermal activity. Moreover, we also recommend comparing thermal anomalies across multiple locations with similar elevations; if a systematic correlation between anomaly magnitude and altitude emerges, it may indicate the presence of a non-volcanic, elevation-related artifact. In all cases, SSTAR should not be used as a black box and the optimal choice of reference area must be evaluated by the user based on the specific volcanic setting, available auxiliary data, and the overall objective of the analysis. Given the importance of selecting an appropriate reference area, we suggest performing sensitivity tests that explore multiple reference configurations; for example, by varying the size of the annular-square layout or testing alternative setups (Fig. 11). Comparing anomaly time series derived from different reference areas provides a practical means of assessing how the ROI behaves relative to various baselines, improves transparency in the interpretation of diffuse thermal anomalies, and helps identify methodological complexities and limitations specific to each case study. When feasible, and particularly when SSTAR is intended for near–real-time monitoring, we recommend (i) computing anomalies simultaneously with respect to multiple reference areas and (ii) adopting a classical training–testing framework commonly used in machine learning and data science (e.g., Girona and Drymoni 2024). In this approach, a configuration is first developed and validated using past events, and its performance is then evaluated on future ones. For example, our results for Shishaldin, validated across four different eruptive episodes, suggest that the selected ROI–reference configuration provides a solid setup for ongoing, near–real-time monitoring of diffuse thermal unrest during subsequent pre-eruptive periods. Conclusions In this manuscript, we introduced SSTAR, a computational tool specifically designed to detect, analyze, and track diffuse, low-temperature thermal anomalies in volcanic and hydrothermal environments using satellite infrared data. These subtle anomalies, typically on the order of 1 K and distributed across several square kilometers, have gained increasing attention as evidence grows that they reflect underlying volcanic unrest and can precede eruptions on timescales of months to years. Because isolating such anomalies is inherently challenging, SSTAR provides a robust, transparent, and operationally oriented workflow for processing long-term thermal infrared time series. Its flexible configuration options, reproducible processing chain, and user-friendly interface enable systematic exploration of diffuse thermal unrest, a class of signals that cannot be adequately characterized using standard hotspot-based algorithms. Together, these capabilities position SSTAR as a methodological advance for research applications, and to facilitate broader adoption, we make the scripts, user-friendly interface, and example datasets openly available. We evaluated SSTAR using Shishaldin volcano (Alaska) as a case study, demonstrating its ability to reveal spatially coherent, pre-eruptive thermal anomalies that emerged well before multiple eruptive events. SSTAR also establishes a foundation for future development by offering a flexible framework that can be extended to incorporate multi-sensor datasets and integrated into multi-hazard monitoring platforms. Building on this capability, SSTAR has been implemented as an online platform at Geosciences Barcelona-CSIC (https://sstar.geo3bcn.csic.es/), where diffuse thermal anomaly results are delivered in near real time for target volcanoes, with the goal of supporting operational monitoring and assisting surveillance agencies and expert committees in alert-level assessments. Supplementary Information Supplementary Material 1.Supplementary Material 2. Supplementary Material 1. Supplementary Material 2.