Magnifying change: Rapid burn scar mapping with multi-resolution, multi-source satellite imagery
Maria Sdraka, Dimitrios Michail, Ioannis Papoutsis · 2026
Read Magnifying change: Rapid burn scar mapping with multi-resolution, multi-source satellite imagery by Maria Sdraka, Dimitrios Michail, Ioannis Papoutsis (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-09262.txt.
Introduction
Maria Sdraka, Dimitrios Michail, and Ioannis Papoutsis map wild fire scars fast from mixed space shots. A scar map guides crews, aid, and loss counts in the days past a fire. Deep tools map well when fine shots are at hand. Fine shots come back slow. Wide shots come fast, but each pixel spans a large plot, so scar edges blur.
The team joins shots of more than one grade and source. Fine facts guide the map where they exist. Wide facts fill gaps when time is short. A reader in fire work gets a map plan built for the first days, when speed can top fine scale in worth.
Catalog record
| CREATOR | Maria Sdraka, Dimitrios Michail, Ioannis Papoutsis |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-09262
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Magnifying change: Rapid burn scar mapping with multi-resolution, multi-source satellite imagery Maria Sdraka, Dimitrios Michail, Ioannis Papoutsis Abstract Delineating wildfire affected areas using satellite imagery remains challenging due to irregular and spatially heterogeneous spectral changes across the electromagnetic spectrum. While recent deep learning approaches achieve high accuracy when high-resolution multispectral data are available, their applicability in operational settings, where a quick delineation of the burn scar shortly after a wildfire incident is required, is limited by the trade-off between spatial resolution and temporal revisit frequency of current satellite systems. To address this limitation, we propose a novel deep learning model, namely BAM-MRCD, which employs multi-resolution, multi-source satellite imagery (MODIS and Sentinel-2) for the timely production of detailed burnt area maps with high spatial and temporal resolution. Our model manages to detect even small scale wildfires with high accuracy, surpassing similar change detection models as well as solid baselines. All data and code are available in the GitHub repository: https://github.com/Orion-AI-Lab/BAM-MRCD . I Introduction Wildfires have become an increasingly global phenomenon, with devastating and long-lasting effects on ecosystems and communities. Human-induced climate change increases the frequency and intensity of weather conditions that facilitate the presence of wildfires — combinations of high temperatures, low humidity and precipitation, and often high winds — resulting in longer fire seasons and a higher risk of extreme wildfire events [ 6 , 14 ] . For example, heat-induced fire weather types in the Mediterranean Basin are projected to escalate by the end of the century [ 29 , 9 , 41 , 20 , 13 ] , posing an imminent threat to already heavily impacted regions. Located in the northeast region of the Mediterranean basin, Greece has been suffering from devastating wildfires over the past two decades [ 16 , 26 ] and studies show that the risk of wildfire in the Greek terrain is bound to increase in the future [ 28 , 23 ] . Therefore, a timely assessment of the damage inflicted by wildfires is of utmost importance for the immediate planning of response and recovery efforts, in order to secure the effective restoration of the ecosystem. Fig. 1 : Outline of our approach. A wildfire event starts at timestep t_{s} and ends at t_{e} . Due to the inherent limitations of the satellite missions, we have daily MODIS captions over the area and sparser Sentinel-2 captions. The proposed method takes advantage of the first available MODIS imagery after the fire ends ( t_{e+1} ), as well as a pair of MODIS and Sentinel-2 captions on the same date, anytime before the fire starts (e.g. t_{s-2} ), and produces a high resolution delineation of the burn scar. This approach alleviates the need for bitemporal high resolution imagery, thus obtaining timely and accurate burn scar maps soon after a wildfire is deemed complete. Traditional ground-based mapping techniques are often limited by accessibility, time constraints, and resource availability. Satellite remote sensing has emerged as a critical tool for post-wildfire assessment, offering a means to quickly and efficiently delineate burnt areas across vast and often inaccessible regions. However, satellite sensors are characterized by distinct spectral, spatial and temporal resolutions, whereas technical limitations usually impose a trade-off between these three dimensions. In an operational setting where prompt assessment of the inflicted damage is a hard requirement for immediate relief efforts, a wildfire monitoring system should be able to leverage the advantages offered by different satellite missions to produce a delineation of the burn scar with the highest level of detail as soon as possible. Among the various satellite sensors available for public use, the Moderate Resolution Imaging Spectroradiometer (MODIS) by NASA and Sentinel-2 by ESA demonstrate a strong potential for this task, each with their own strengths and limitations. MODIS provides frequent observations at moderate spatial resolution, allowing near real-time monitoring of fire impact, while Sentinel-2 offers high spatial resolution imagery that enhances the accuracy of the burn scar perimeter delineation, but with a lower revisit frequency spanning over a few days. This study presents a novel technique which integrates data from these two sensors to improve the timeliness and precision of burnt area mapping, employing the FLOGA dataset [ 30 ] which provides high-resolution manual annotations for various wildfire events in Greece accompanied by multi-source bitemporal satellite imagery. The proposed approach leverages the complementary strengths of the Sentinel-2 and MODIS satellites, i.e. the high temporal resolution of MODIS and the superior spatial resolution of Sentinel-2, to generate accurate burn scar maps shortly after a wildfire (Fig. 1 ). By utilizing a custom deep learning architecture, our method aims to provide an efficient and reliable solution for timely wildfire impact assessment, capable of handling burn scars of varying size. We hope that our work will propel further research in this newly emerged field and contribute to the next-day planning of relief and ecosystem restoration efforts. In short, our contributions are summarized below: • We formulate a Burnt Area Mapping (BAM) problem with multi-resolution satellite input for change detection, which is distinct from generic change detection tasks. • We propose a new deep learning approach, namely BAM-MRCD, which exploits multi-scale feature extraction from pre-fire Sentinel-2 and bitemporal MODIS imagery. Our model processes both input sources in a parallel fashion, without the need to use super-resolution modules which may potentially impute noise to the pipeline. • BAM-MRCD outperforms the baselines as well as other models for multi-resolution change detection proposed in the literature, achieving high accuracy on wildfire impacted areas of varying size. II Related work II-A Change detection with multi-resolution input Change detection (CD) tasks typically require a bitemporal input which comprises one image before and one image after the change under investigation. Time restrictions surfacing in operational emergency settings, such as during or shortly after natural disasters, often dictate the use of different satellite sources in order to obtain an initial assessment of the damage as soon as possible. However, this can lead to pairs of images with vastly different spatial and/or spectral resolution, thus hindering the accuracy of traditional change detection algorithms that are based on similar input pairs. To tackle this issue, various change detection approaches utilising multi-resolution satellite imagery have been proposed in the literature over the past years. A number of these studies employ a two-step approach, combining a super-resolution (SR) module for the enhancement of the low-resolution image [ 31 ] , along with a separate change detection module for the computation of the final output. For example, RACDNet [ 34 ] comprises a WDSR model [ 39 ] enhanced with corner detection which is initially trained for the SR task, and a siamese UNet [ 27 ] trained for the CD task. In a similar fashion, MF-SRCDNet [ 18 ] employs a Res-UNet [ 2 ] to downscale the low resolution imagery and a modified STANet [ 3 ] for the production of the binary change map which takes as input an auxiliary layer with edge information. Other studies propose an end-to-end approach where both SR and CD modules are trained simultaneously through a carefully designed compound loss, such as SRCDNet [ 22 ] and ESR-DMNet [ 19 ] . Finally, a hybrid approach is followed by DCILNet [ 17 ] ,