FireCastNet: earth-as-a-graph for seasonal fire prediction Authors: Dimitrios Michail, Charalampos Davalas, Konstantinos Chafis, Lefki-Ioanna Panagiotou, Ioannis Prapas, Nikolaos Ioannis Bountos, Spyros Kondylatos, Ioannis Papoutsis Published: 2025 ABSTRACT With climate change intensifying fire weather conditions globally, accurate seasonal wildfire forecasting has become critical for disaster preparedness and ecosystem management. We introduce FireCastNet, a novel deep learning architecture that combines 3D convolutional encoding with GraphCast-based Graph Neural Networks (GNNs) to model complex spatio-temporal dependencies for global wildfire prediction. Our approach leverages the SeasFire dataset, a comprehensive multivariate Earth system datacube containing climate, vegetation, and human-related variables, to forecast burned area patterns up to 6 months in advance. FireCastNet treats the Earth as an interconnected graph, enabling it to capture both local fire dynamics and long-range teleconnections that influence wildfire behavior across different spatial and temporal scales. Through comprehensive benchmarking against state-of-the-art models including GRU, Conv-GRU, Conv-LSTM, U-TAE, and TeleViT, we demonstrate that FireCastNet achieves superior performance in global burned area forecasting, with particularly strong results in fire-prone regions such as Africa, South America, and Southeast Asia. Our analysis reveals that longer input time-series significantly improve prediction robustness, while spatial context integration enhances model performance across extended forecasting horizons. Additionally, we implement local area modeling techniques that provide enhanced spatial resolution and accuracy for region-specific predictions. These findings highlight the importance of modeling Earth system interactions for long-term wildfire prediction. Introduction Fire plays a pivotal role in the Earth system, significantly influencing ecosystems worldwide. While traditionally viewed as a carbon-neutral process over the long term, climate change disrupts this balance through the intensification of fire weather conditions, leading to a rise in global fire activity1. A feedback loop is created when fires encroach into evergreen forest regions, posing a threat to their role as carbon sinks. This situation triggers the release of stored carbon into the atmosphere, further exacerbating climate change2. Moreover, wildfires represent a critical natural hazard with far-reaching consequences for societies worldwide, resulting in loss of life, property, infrastructure, and ecosystem services3. Hence, it is imperative to improve our understanding and forecasting capabilities of wildfire phenomena within the Earth system. By doing so, we can formulate effective strategies to mitigate the adverse impacts of wildfires and climate change. These strategies may encompass enhanced forestry management, strengthened infrastructure resilience, disaster preparedness, and the implementation of more accurate warning systems. Producing reliable long-range wildfire forecasts, spanning up to 6 months, requires advanced techniques that account for the Earth as one interconnected system. This system is shaped by continuous interactions across large spatio-temporal scales. These interactions manifest in two key ways: (a) memory effects, where past conditions like fuel accumulation and drought influence future fire activity, and (b) teleconnections, where distant climate events and weather patterns affect wildfire risk. Traditional statistical models, which rely on historical data from the same time period, have been widely used for fire hazard forecasting. However, these models often struggle to capture the intricate, non-linear patterns and dependencies that govern wildfire behavior. Given the critical importance of predicting fire occurrences to protect ecosystems and human life, it is essential to recognize and model these complex interactions effectively. In this study, we present FireCastNet, a novel architecture aimed at enhancing seasonal wildfire prediction by utilizing the GraphCast4 weather forecasting model. FireCastNet aims at leveraging both short- and long-range dependencies across the Earth system and providing a deeper and more comprehensive understanding of the spatio-temporal interactions that influence wildfire behavior. Unlike traditional models, FireCastNet does not rely on proxy variables like pre-encoded oceanic indices for temporal dynamics. Instead, it directly models these dependencies from raw time series data, offering a more flexible and data-driven approach to long-range fire forecasting. We conduct a thorough experimental evaluation, benchmarking FireCastNet’s performance against baseline methods, such as the seasonal mean cycle, and several state-of-the-art machine learning models, including Gated Recurrent Units (GRU), Convolutional GRU (Conv-GRU), Convolutional Long Short-Term Memory (Conv-LSTM), and the U-Net Transformer-based Attention Encoder (U-TAE). This evaluation explores the predictive capabilities of each model across various spatio-temporal contexts, assessing factors such as different time-series lengths, spatial neighborhood sizes, and prediction horizons. Our experiments focus on forecasting burned area patterns weeks to months in advance, at a global scale and with a spatial resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.25^{\circ }$$\end{document}. By leveraging spatio-temporal data at multiple scales, we assess how effectively each deep learning model captures the intricate dynamics of wildfire behavior. To enhance the spatial resolution and accuracy of our predictions, we employ local area modeling approaches4,5. These techniques are integrated into FireCastNet to enable more precise localized forecasting. Our evaluation strategy involves a two-tier approach: first, we assess global models across regions, which are defined by the Global Fire Emissions Database6 (GFED) and encompass different biomes and climatic zones. Second, we leverage local area modeling by training region-specific models for each GFED region, thereby further improving performance. This methodology allows us to evaluate the model’s performance across diverse geographical contexts, providing insights into its adaptability and effectiveness in various fire-prone regions. The main contributions of this work can be summarized as follows:Introduction of FireCastNet, a deep learning architecture that combines 3D convolutional encoding with GraphCast-based Graph Neural Networks (GNNs) to model spatio-temporal dependencies for global, seasonal wildfire prediction.Leveraging the open-access multivariate SeasFire Earth system datacube, FireCastNet forecasts global burned area patterns up to six months in advance, demonstrating that incorporating longer input sequences and spatial context improves predictive robustness compared to traditional and state-of-the-art machine learning baselines (GRU, Conv-GRU, Conv-LSTM, U-TAE, and TeleViT).We integrate region-specific local modeling approaches4,5 within FireCastNet, enabling higher-resolution and more accurate predictions across diverse GFED-defined biomes.All source code can be found at https://github.com/seasfire/firecastnet. Introduction of FireCastNet, a deep learning architecture that combines 3D convolutional encoding with GraphCast-based Graph Neural Networks (GNNs) to model spatio-temporal dependencies for global, seasonal wildfire prediction. Introduction of FireCastNet, a deep learning architecture that combines 3D convolutional encoding with GraphCast-based Graph Neural Networks (GNNs) to model spatio-temporal dependencies for global, seasonal wildfire prediction. Leveraging the open-access multivariate SeasFire Earth system datacube, FireCastNet forecasts global burned area patterns up to six months in advance, demonstrating that incorporating longer input sequences and spatial context improves predictive robustness compared to traditional and state-of-the-art machine learning baselines (GRU, Conv-GRU, Conv-LSTM, U-TAE, and TeleViT). Leveraging the open-access multivariate SeasFire Earth system datacube, FireCastNet forecasts global burned area patterns up to six months in advance, demonstrating that incorporating longer input sequences and spatial context improves predictive robustness compared to traditional and state-of-the-art machine learning baselines (GRU, Conv-GRU, Conv-LSTM, U-TAE, and TeleViT). We integrate region-specific local modeling approaches4,5 within FireCastNet, enabling higher-resolution and more accurate predictions across diverse GFED-defined biomes. We integrate region-specific local modeling approaches4,5 within FireCastNet, enabling higher-resolution and more accurate predictions across diverse GFED-defined biomes. Related work Wildfires are notoriously hard to model due to the non-linear interactions between the different Earth system processes that affect them7. Weather, vegetation and humans interact in evolving ways that contribute in the expansion or suppression of wildfires. Reichstein et al.8 propose Deep Learning as a method to learn in a data-driven way these complex spatio-temporal interactions that influence wildfires. Several recent studies have used deep learning for wildfire-related use cases9. For short-term daily predictions the temporal context is mostly enough, with spatio-temporal models offering little to no advantage10,11. For longer-term predictions, i.e. on subseasonal to seasonal scales, very few works have studied the effect of spatial and temporal context12–14. Joshi et al.12 use monthly aggregated input to predict burned area using multi-layer neural networks. Li et al.15 includes the temporal aspect of the input using a temporal attention network on time-series Recent studies rely on the SeasFire Datacube16 to train models with seasonal forecasting skill. Prapas et al.13 use temporal snapshots of the fire drivers to predict future burned area patterns, defining burned area forecasting as a segmentation task and demonstrating skillful forecasts even 2 months in advance. An expansion of this14, recognises the need to view the earth as a system for long-term forecasting and proposes a novel architecture that leverages teleconnection indices and global input in conjunction with those snapshots. This setup helps improve long-term skill, but ignores the temporal component of the input variables. Zhao et al.17 integrate causality with GNNs to explicitly model the causal mechanism among complex variables via graph learning, and test their models in the European boreal and Mediterranean biome. Finally, in the work of Zhu et al.18 the authors expand on the TeleViT architecture and add a balancing term to handle the data imbalance between the burned and non-burned classes. This allows them to create a fire risk index that is better calibrated between different regions. Machine learning has recently made significant advances in global weather forecasting19, with graph neural networks showing tremendous potential to represent the Earth as a system20,21. Keisler20 introduces a GNN-based model that predicts multiple atmospheric variables across various pressure levels on a global scale. Building upon this foundation, Lam et al.21 introduced the GraphCast model, a machine learning method that forecasts hundreds of weather variables over a 10-day period at 0.25-degree resolution globally, delivering results in under a minute. Further advancing the GraphCast architecture, the work of Oskarsson et al.22 introduces Graph-EFM, a probabilistic weather forecasting model that combines a flexible latent-variable formulation with a hierarchical graph-based framework. This model efficiently generates spatially coherent ensemble forecasts, achieving errors equivalent to or lower than comparable deterministic models while accurately capturing forecast uncertainty. These types of models have also been shown to be effective at the temporal scales that we are interested in this study, with Fuxi-S2S23 achieving skillfull forecasts at the subseasonal to seasonal scales. Our work aims to leverage the advances in data-driven weather modeling for burned area forecasting with an architecture inspired by Graphcast. Complementing global approaches, the paradigm of Local Area Modeling has emerged as a powerful strategy for improving forecast skill in limited geographic regions. Oskarsson et al.5 propose a graph-based neural weather prediction method that tailors modeling to limited areas rather than the entire globe. This approach focuses computational resources locally, enabling higher resolution and potentially finer-grained feature learning in regions of interest, which is particularly beneficial when computational constraints or application domain focus (e.g., wildfire-prone areas) restrict the scope of forecasts. Seasonal fire forecasting Problem formulation We view the problem as binary classification at a particular location of the cube and a particular timestamp. Thus, given a triplet \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\phi _c, \lambda _c, t_c)$$\end{document} of latitude, longitude and 8-day period we predict whether a fire will occur or not at that particular location in time and space. As input we use a timeseries for each variable of length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ts \in \{6, 12, 24\}$$\end{document} timesteps, where each timestep corresponds to a single 8-day period. As target variable we predict the presence of burned areas at a future timestep \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t+h$$\end{document} with different values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h \in \{1, 2, 4, 8, 16, 24\}$$\end{document}. Dataset Table 1Fire driver variables (10 input, 1 static and 2 positional) used for predictions.Input typeVariableAbbreviationBasic inputMean sea level pressuremslpTotal precipitationtpVapor pressure deficitvpdSea surface temperaturesstTemperature at 2 m - Meant2m_meanSurface solar radiation downwardsssrdVolumetric soil water level 1swvl1Land surface temperature at daylst_dayNormalized difference vegetation indexndviPopulation densitypop_densStatic inputLand sea masklsmPositional inputLatitude (sin/cos)latLongitude (sin/cos)lonTargetBurned areas (as binary)gwis_baTarget variable represents burned area in hectares (converted to binary for classification tasks). Table 1Fire driver variables (10 input, 1 static and 2 positional) used for predictions.Input typeVariableAbbreviationBasic inputMean sea level pressuremslpTotal precipitationtpVapor pressure deficitvpdSea surface temperaturesstTemperature at 2 m - Meant2m_meanSurface solar radiation downwardsssrdVolumetric soil water level 1swvl1Land surface temperature at daylst_dayNormalized difference vegetation indexndviPopulation densitypop_densStatic inputLand sea masklsmPositional inputLatitude (sin/cos)latLongitude (sin/cos)lonTargetBurned areas (as binary)gwis_baTarget variable represents burned area in hectares (converted to binary for classification tasks). Fire driver variables (10 input, 1 static and 2 positional) used for predictions. Target variable represents burned area in hectares (converted to binary for classification tasks). The SeasFire Datacube16 is an analysis-ready, open-access datacube fit for wildfire forecasting at different spatio-temporal scales that serves not only for seasonal fire forecasting worldwide but also for forecasting emissions from wildfires and predicting the evolution of wildfire regimes. It spans over 21 years (2001–2021) with an 8-day temporal and a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.25^{\circ }$$\end{document} spatial resolution. The dataset provides a comprehensive coverage of atmospheric, climatological, vegetation and socioeconomic factors influencing wildfires and contains 58 variables in total, along with target variables, such as burned areas, fire radiative power, and carbon emissions from wildfires. 11 different variables were utilized (see Table 1) in order to predict the existence of fire in different time horizons. In addition and when needed, we also used simple positional encodings by augmenting the feature vector with the actual cube coordinates, i.e. sin/cos of latitude and longitude. All variables, except the cube coordinates, were standardized before use. Naive forecasting When dealing with seasonal time series, the naive seasonal forecasting method uses the observed values in the same period of the previous seasons24. For example in order to predict for a particular week of August in a particular year, one could use the observed value in the same week of August one year before. Our dataset contains multiple years, allowing for different baseline methods based on the utilization of past data. We utilize the following two:To forecast the occurrence of fire at a particular location in a particular 8-day period of the year, we examine whether there were fires at that location in the specific 8-day period in any of the previous years. If there were, the model predicts a fire; otherwise, it predicts no fire. We call this baseline the any-fire baseline.To forecast the occurrence of fire at a particular location in a particular 8-day period of the year we use a majority rule. Thus, we predict fire if the number of previous years with fire in that particular 8-day period is larger than the number of previous years without fire at that period. This baseline is called the majority baseline.The majority baseline approach aligns effectively with the binary classification task. To forecast the occurrence of fire at a particular location in a particular 8-day period of the year, we examine whether there were fires at that location in the specific 8-day period in any of the previous years. If there were, the model predicts a fire; otherwise, it predicts no fire. We call this baseline the any-fire baseline. To forecast the occurrence of fire at a particular location in a particular 8-day period of the year, we examine whether there were fires at that location in the specific 8-day period in any of the previous years. If there were, the model predicts a fire; otherwise, it predicts no fire. We call this baseline the any-fire baseline. To forecast the occurrence of fire at a particular location in a particular 8-day period of the year we use a majority rule. Thus, we predict fire if the number of previous years with fire in that particular 8-day period is larger than the number of previous years without fire at that period. This baseline is called the majority baseline. To forecast the occurrence of fire at a particular location in a particular 8-day period of the year we use a majority rule. Thus, we predict fire if the number of previous years with fire in that particular 8-day period is larger than the number of previous years without fire at that period. This baseline is called the majority baseline. Baseline models Gated recurrent units Recurrent Neural Networks25 are a natural choice when trying to capture temporal dependencies. We utilize a very simple architecture comprising of Gated Recurrent Units (GRU). The neural network is comprised of two GRU layers with 64 hidden channels in each layer. A dropout layer with a probability of 0.1 exists between the two GRU layers. As the task is binary classification, a final linear layer reduces the representation into a single prediction and a sigmoid function outputs the final prediction. Convolutional GRU To capture both temporal and spatial dependencies at the same time, we utilize a convolutional GRU network26 (Conv-GRU). In such a network all inputs, cell outputs, hidden states and gates are 3D tensors where the last two dimensions are spatial dimensions (rows and columns). This architecture is particularly effective for spatio-temporal tasks, as it incorporates convolution operations in the update and reset gates to model spatial relationships. The Conv-GRU model is defined using the following set of equations, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$*$$\end{document} denotes the convolution operator, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\odot$$\end{document} is element-wise multiplication, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document} is the sigmoid activation function:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \varvec{\textsf{z}}_t&= \sigma (\textbf{W}_{xz} * \varvec{\textsf{x}}_t + \textbf{W}_{hz} * \varvec{\textsf{h}}_{t-1} + \varvec{\textsf{b}}_z), \\ \varvec{\textsf{r}}_t&= \sigma (\textbf{W}_{xr} * \varvec{\textsf{x}}_t + \textbf{W}_{hr} * \varvec{\textsf{h}}_{t-1} + \varvec{\textsf{b}}_r), \\ \tilde{\varvec{\textsf{h}}}_t&= \tanh (\textbf{W}_{x\tilde{h}} * \varvec{\textsf{x}}_t + \varvec{\textsf{r}}_t \odot (\textbf{W}_{h\tilde{h}} * \varvec{\textsf{h}}_{t-1}) + \varvec{\textsf{b}}_{\tilde{h}}), \\ \varvec{\textsf{h}}_t&= (1 - \varvec{\textsf{z}}_t) \odot \varvec{\textsf{h}}_{t-1} + \varvec{\textsf{z}}_t \odot \tilde{\varvec{\textsf{h}}}_t. \end{aligned}$$\end{document}Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}_t \in \mathbb {R}^{F \times H \times W}$$\end{document} represents the input tensor at time step \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F$$\end{document} is the number of features, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W$$\end{document} are the spatial dimensions. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{h}}_{t-1}$$\end{document} is the hidden state from the previous time step, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{z}}_t$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{r}}_t$$\end{document} are the update and reset gates, respectively. The hidden state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{h}}_t$$\end{document} is updated by controlling how much of the past hidden state is carried forward (via \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{z}}_t$$\end{document}) and how much of the new information (via \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\varvec{\textsf{h}}}_t$$\end{document}) is incorporated into the current state. Given \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\phi _c, \lambda _c)$$\end{document} for a particular location of interest and a particular time t we generate a 3D tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}_t$$\end{document} whose last two dimensions are the number of rows and columns of a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(2r+1) \times (2r+1)$$\end{document} spatial grid. The first dimension is the number of features. The grid is centered at our location of interest while the remaining locations represent \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\phi _c \pm i \cdot 0.25^{\circ }, \lambda _c \pm i \cdot 0.25^{\circ })$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i \in \{1, \ldots , r-1\}$$\end{document}. We feed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}_t$$\end{document} to the model for each timestep t in order to compute the hidden representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{h}}_t$$\end{document}. The final result is obtained by utilizing multilayer perceptron (MLP) layers in order to gradually aggregate the information into a single prediction. Convolutional LSTM We also utilize a convolutional LSTM network27 (Conv-LSTM). Again all inputs, cell outputs, hidden states and gates are 3D tensors where the last two dimensions are spatial dimensions (rows and columns). The future state of a certain cell is computed from the input and the past states of its local neighbors. This is achieved by using convolution operators in different transitions such as the state-to-state and/or input-to-state. We utilize a slightly different network from the one in Shi et al.27, which is implemented using the following equations:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \varvec{\textsf{i}}_t&= \sigma (\textbf{W}_{xi} * \varvec{\textsf{x}}_t + \textbf{W}_{hi} * \varvec{\textsf{h}}_{t-1} + \varvec{\textsf{b}}_i)\\ \varvec{\textsf{f}}_t&= \sigma (\textbf{W}_{xf} * \varvec{\textsf{x}}_t + \textbf{W}_{hf} * \varvec{\textsf{h}}_{t-1} + \varvec{\textsf{b}}_f)\\ \varvec{\textsf{o}}_t&= \sigma (\textbf{W}_{xo} * \varvec{\textsf{x}}_t + \textbf{W}_{ho} * \varvec{\textsf{h}}_{t-1} + \varvec{\textsf{b}}_o)\\ \varvec{\textsf{g}}_t&= \tanh (\textbf{W}_{xg} * \varvec{\textsf{x}}_t + \textbf{W}_{hg} * \varvec{\textsf{h}}_{t-1} + \varvec{\textsf{b}}_g)\\ \varvec{\textsf{c}}_t&= \varvec{\textsf{f}}_t \odot \varvec{\textsf{c}}_{t-1} + \varvec{\textsf{i}}_t \odot \varvec{\textsf{g}}_t\\ \varvec{\textsf{h}}_t&= \varvec{\textsf{o}}_t \odot \tanh (\varvec{\textsf{c}}_t). \end{aligned}$$\end{document}Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$*$$\end{document} denotes the convolution operator and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\odot$$\end{document} point-wise multiplication. We integrate the Conv-LSTM into the classification pipeline the same way as the Conv-GRU architecture. U-Net with temporal attention encoder The U-Net with Temporal Attention Encoder (U-TAE)28 architecture is designed for spatio-temporal feature extraction and segmentation of satellite image time series. It extends the classic U-Net by incorporating a temporal self-attention mechanism to process temporal sequences of images. This enables the network to capture temporal dependencies across multiple resolution levels. The U-TAE consists of two main components: the spatial encoder–decoder network and the temporal self-attention module. The spatial encoder–decoder network follows the U-Net architecture and processes each image in the time series separately. The encoder applies convolutional layers to extract spatial features from the images at different resolutions. Specifically, each encoder block consists of two \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3 \times 3$$\end{document} convolutions followed by Group Normalization and ReLU activations, halving the spatial resolution at each level. The decoder performs upsampling using transposed convolutions, and the corresponding encoder feature maps are concatenated at each level (U-Net-style skip connections). To capture the temporal dynamics, U-TAE incorporates a Lightweight-Temporal Attention Encoder29 (L-TAE) module, which is applied independently to each pixel’s feature map sequence extracted from the lowest-resolution level of the spatial encoder. The L-TAE computes temporal attention masks over the sequence, allowing the model to focus on relevant time steps. These temporal attention maps are then bilinearly interpolated to match the resolutions of the higher levels of the encoder. After the attention is applied to each feature map group, the feature maps are concatenated and passed through a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 1$$\end{document} convolution to fuse temporal information. FireCastNet Fig. 1High-level representation of our architecture. The FireCastNet architecture consists of three main components: (a) a cube embedding block for spatio-temporal feature extraction using a 3D convolution layer that reduces the spatial and temporal dimensions of the input timeseries data (b) a GraphCast block, which leverages a multi-mesh Graph Neural Network (GNN) to model long-range spatial interactions across a refined icosahedral mesh structure, and (c) a sub-pixel convolution block for upscaling the output to match the resolution of the original input. Fig. 2Target variable and prediction (per horizon) of the best global model, using a time-series of length 24, on 30 September 2019. The target variable is always the same. The input variables depend on the prediction horizon, i.e. for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h=1$$\end{document} we use a time-series of length 24 ending one 8-day period before, for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h=2$$\end{document} we use a time-series of length 24 ending two 8-day periods before, and so on. The sigmoid output represents the prediction confidence. Figure generated using the Matplotlib30 library (https://zenodo.org/records/17595503). High-level representation of our architecture. The FireCastNet architecture consists of three main components: (a) a cube embedding block for spatio-temporal feature extraction using a 3D convolution layer that reduces the spatial and temporal dimensions of the input timeseries data (b) a GraphCast block, which leverages a multi-mesh Graph Neural Network (GNN) to model long-range spatial interactions across a refined icosahedral mesh structure, and (c) a sub-pixel convolution block for upscaling the output to match the resolution of the original input. Target variable and prediction (per horizon) of the best global model, using a time-series of length 24, on 30 September 2019. The target variable is always the same. The input variables depend on the prediction horizon, i.e. for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h=1$$\end{document} we use a time-series of length 24 ending one 8-day period before, for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h=2$$\end{document} we use a time-series of length 24 ending two 8-day periods before, and so on. The sigmoid output represents the prediction confidence. Figure generated using the Matplotlib30 library (https://zenodo.org/records/17595503). The FireCastNet architecture is shown in Fig. 1. It consists of a cube embedding block, a GraphCast block and a sub-pixel convolution for upsampling. This architecture is designed to perform segmentation tasks on a timeseries of multi-channel images. The input data is represented as a tensor of dimensions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}} \in \mathbb {R}^{T \times C \times H \times W}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T$$\end{document} is the number of temporal steps, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C$$\end{document} is the number of input channels, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W$$\end{document} are the height and width of each image, respectively. For reference, ERA5 has spatial dimensions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$721 \times 1440$$\end{document}, while the SeasFire datacube uses \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$720 \times 1440$$\end{document}. Thus, in our case, we have \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H = 720$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W = 1440$$\end{document}. Cube embedding The first stage of FireCastNet is a 3D convolutional block designed for spatio-temporal feature extraction. Its space–time cube-embedding reduces the spatial and temporal dimensions of the input, thereby accelerating training31. This embedding applies a 3D convolution with kernel and stride \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T \times \frac{H}{H'} \times \frac{W}{W'}$$\end{document}, and increases the channel dimension from C to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C' > C$$\end{document}. Formally, the block is expressed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}' = \text {Conv3D}(\varvec{\textsf{x}}; \theta )$$\end{document}, and produces an output tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\mathsf {x'}} \in \mathbb {R}^{C' \times H' \times W'}$$\end{document}. Maintaining a spatial resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.25^\circ$$\end{document} requires \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{H}{H'} = 1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{W}{W'} = 1$$\end{document}. Instead, we reduce the resolution to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1^\circ$$\end{document} by setting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{H}{H'} = 4$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{W}{W'} = 4$$\end{document}. The model uses 11 variables from the datacube (Table 1), along with additional channels containing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\cos (\text {latitude})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sin /\cos (\text {longitude})$$\end{document}, giving \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C = 14$$\end{document}. We expand this to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C' = 64$$\end{document}. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H'=180$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W'=360$$\end{document}, the resulting cube-embedding output is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\mathsf {x'}} \in \mathbb {R}^{64 \times 180 \times 360}$$\end{document}. GraphCast The core of the FireCastNet architecture is GraphCast, which uses a Graph Neural Network (GNN) in an “encode-process-decode” configuration. To capitalize on the GNN’s ability to handle arbitrary sparse interactions, it uses an internal “multi-mesh” graph representation. This structure facilitates the modeling of long-range interactions within a few message-passing steps and provides a more homogeneous spatial resolution across the globe compared to a standard latitude–longitude grid. The latitude–longitude grid is a less suitable representation due to its spatial inhomogeneity and the high density of grid points at the poles, which demands disproportionate computational resources. The multi-mesh is a spatially homogeneous graph with high spatial resolution over the globe. It is constructed by iteratively refining a regular icosahedron (a polyhedron with 12 nodes, 20 faces, and 30 edges) six times. In each refinement step, every triangular face is subdivided into four smaller triangles by connecting the midpoints of its edges, leading to four times more faces and edges. The newly created vertices are then reprojected onto a circumscribing sphere to ensure the mesh closely approximates the shape of the Earth. This hierarchical process yields a series of increasingly finer meshes. The multi-mesh ultimately contains the 40,962 nodes from the highest resolution mesh, and the union of all the edges created in the intermediate graphs, forming a flat hierarchy of edges with varying lengths. This multi-scale graph allows coarse edges to bridge long distances and fine edges to capture local interactions. The module operates in three stages:Encoder The encoder maps the input data from the latitude–longitude grid to the multi-mesh. It uses a GNN with directed edges from the grid points to the multi-mesh nodes to encode the grid-based features into learned node attributes. At this stage, additional positional information for the mesh nodes and edges is also encoded along with the features from the grid. The output is an initial set of node embeddings on the multi-mesh, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{h}}_0 \in \mathbb {R}^{|V| \times C_{\text {mesh}}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|V|$$\end{document} is the number of mesh nodes. We use \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{\text {mesh}} = 64$$\end{document}.Processor The processor consists of 12 consecutive GNN layers that perform learned message-passing on the multi-mesh graph. Each layer updates the node embeddings, allowing information to propagate efficiently across space. The long-range edges in the multi-scale mesh structure are crucial for capturing distant teleconnections. The update at each layer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l$$\end{document} can be represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{h}}_{l+1} = \text {GNN}_{\text {process}}(\varvec{\textsf{h}}_l, G)$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$G$$\end{document} is the multi-mesh graph.Decoder Finally, the decoder maps the processed node embeddings from the multi-mesh back to the latitude–longitude grid. It employs a GNN with directed edges from the multi-mesh nodes to the grid points. This grid representation is then combined with the input state to form the final output prediction. The decoding step is represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}_{\text {dec}} = \text {GNN}_{\text {dec}}(\varvec{\textsf{h}}_{12}; G)$$\end{document}. Encoder The encoder maps the input data from the latitude–longitude grid to the multi-mesh. It uses a GNN with directed edges from the grid points to the multi-mesh nodes to encode the grid-based features into learned node attributes. At this stage, additional positional information for the mesh nodes and edges is also encoded along with the features from the grid. The output is an initial set of node embeddings on the multi-mesh, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{h}}_0 \in \mathbb {R}^{|V| \times C_{\text {mesh}}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|V|$$\end{document} is the number of mesh nodes. We use \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{\text {mesh}} = 64$$\end{document}. Encoder The encoder maps the input data from the latitude–longitude grid to the multi-mesh. It uses a GNN with directed edges from the grid points to the multi-mesh nodes to encode the grid-based features into learned node attributes. At this stage, additional positional information for the mesh nodes and edges is also encoded along with the features from the grid. The output is an initial set of node embeddings on the multi-mesh, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{h}}_0 \in \mathbb {R}^{|V| \times C_{\text {mesh}}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|V|$$\end{document} is the number of mesh nodes. We use \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{\text {mesh}} = 64$$\end{document}. Processor The processor consists of 12 consecutive GNN layers that perform learned message-passing on the multi-mesh graph. Each layer updates the node embeddings, allowing information to propagate efficiently across space. The long-range edges in the multi-scale mesh structure are crucial for capturing distant teleconnections. The update at each layer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l$$\end{document} can be represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{h}}_{l+1} = \text {GNN}_{\text {process}}(\varvec{\textsf{h}}_l, G)$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$G$$\end{document} is the multi-mesh graph. Processor The processor consists of 12 consecutive GNN layers that perform learned message-passing on the multi-mesh graph. Each layer updates the node embeddings, allowing information to propagate efficiently across space. The long-range edges in the multi-scale mesh structure are crucial for capturing distant teleconnections. The update at each layer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l$$\end{document} can be represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{h}}_{l+1} = \text {GNN}_{\text {process}}(\varvec{\textsf{h}}_l, G)$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$G$$\end{document} is the multi-mesh graph. Decoder Finally, the decoder maps the processed node embeddings from the multi-mesh back to the latitude–longitude grid. It employs a GNN with directed edges from the multi-mesh nodes to the grid points. This grid representation is then combined with the input state to form the final output prediction. The decoding step is represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}_{\text {dec}} = \text {GNN}_{\text {dec}}(\varvec{\textsf{h}}_{12}; G)$$\end{document}. Decoder Finally, the decoder maps the processed node embeddings from the multi-mesh back to the latitude–longitude grid. It employs a GNN with directed edges from the multi-mesh nodes to the grid points. This grid representation is then combined with the input state to form the final output prediction. The decoding step is represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}_{\text {dec}} = \text {GNN}_{\text {dec}}(\varvec{\textsf{h}}_{12}; G)$$\end{document}. Upscaling The final stage of FireCastNet is a sub-pixel convolution32 block, which upscales the decoded output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}_{\text {dec}}$$\end{document} to match the spatial dimensions of the original input. We choose the number of channels of the decoder to be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r^2$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r$$\end{document} is the upscaling factor (in our case, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r = \frac{H}{H'} = \frac{W}{W'} = 4$$\end{document} to restore the original \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.25^\circ$$\end{document} resolution). Thus, the decoder produces \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}_{\text {dec}} \in \mathbb {R}^{16 \times 180 \times 360}$$\end{document}. The sub-pixel convolution rearranges the feature maps into higher-resolution output, producing the final segmentation map of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H \times W = 720 \times 1440$$\end{document}. The upscaling operation is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{\textsf{x}}_{\text {out}} = \text {Sub-Pixel-Conv}(\varvec{\textsf{x}}_{\text {dec}})$$\end{document}, yielding the segmentation output.Table 2Comparison of all global-scale models.Model/Horizon12481624GRU0.5580.5490.5400.5190.5210.520Conv-GRU0.5930.5790.5560.5510.5450.569Conv-LSTM0.6350.6080.6150.6010.5980.589U-TAE0.6240.6210.6110.6010.6080.594TeleViT140.6220.6170.6100.6110.6060.604FireCastNet (ours)0.6410.6360.6310.6310.6280.633The time-series length is 24 (except for TeleViT, which cannot process time-series). Table 2Comparison of all global-scale models.Model/Horizon12481624GRU0.5580.5490.5400.5190.5210.520Conv-GRU0.5930.5790.5560.5510.5450.569Conv-LSTM0.6350.6080.6150.6010.5980.589U-TAE0.6240.6210.6110.6010.6080.594TeleViT140.6220.6170.6100.6110.6060.604FireCastNet (ours)0.6410.6360.6310.6310.6280.633The time-series length is 24 (except for TeleViT, which cannot process time-series). Comparison of all global-scale models. The time-series length is 24 (except for TeleViT, which cannot process time-series). Experiments Our methodology can be directly applied to any region of the world, producing different models per region. For our forecasting task, our split is time-based, using years 2002-2017, 2018 and 2019 for train, validation, and test, respectively. The SeasFire dataset is utilized in all the experiments (see “Dataset” section). Figure 2 shows the ground truth and model predictions for September 30, 2019. For each time horizon \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h=\{1, 2, 4, 8, 16, 24\}$$\end{document}, we train a separately model. Each such model uses a time-series of length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ts=24$$\end{document} ending h periods before the target date, predicting the presence of fire in the next h 8-day period. Due to our highly imbalanced dataset, the performance of the model is evaluated using the Area Under the Precision-Recall Curve (AUPRC)33. We calculate the metric using the average precision score, which summarizes a precision-recall curve as a weighted mean of precisions at each threshold, with the difference in recall from the previous threshold as weight\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$AP = \underset{t}{\sum }\ (R_t - R_{t-1}) \cdot P_t,$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_t, R_t$$\end{document} is the respective precision and recall at threshold index t.Fig. 3Model performance at a global scale with different time-series length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ts \in \{6, 12, 24\}$$\end{document} periods (8-days). Model performance at a global scale with different time-series length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ts \in \{6, 12, 24\}$$\end{document} periods (8-days). Models are trained for 50 epochs. We use binary cross-entropy loss and the AdamW34 optimizer. The initial learning rate is set to 0.001 and adjusted using SGDR35. The schedule contains two cycles with 10 and 40 epochs, respectively. A weight decay factor of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-7}$$\end{document} is used when training all models. Global model In this section, we train and compare FireCastNet at a global scale. The naive prediction baseline models discussed in “Naive forecasting” section exhibit 0.174 (anyfire) and 0.3211 (majority) AUPRC. The first baseline checks whether any previous year contains some fire in the same period, while the second checks whether a majority of the previous years contains fires. All models take as input the 11 inputs described in Table 1. The GRU model is comprised of a single hidden layer with 64 hidden units. The Conv-GRU and Conv-LSTM models use the same parameters as the GRU with the only difference of an additional setting of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5 \times 5$$\end{document} kernel. The U-TAE model is configured with 16 attention heads, and a layer size of 256. Each attention head has a key dimension of 4. Finally, FireCastNet processes data with spatial resolution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.25^{\circ }$$\end{document} and geographic bounds from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-89.875^{\circ }$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$89.875^{\circ }$$\end{document} latitude and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-179.875^{\circ }$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$179.875^{\circ }$$\end{document} longitude. The Conv3d block decreases spatial resolution to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1^{\circ }$$\end{document}, and adjusts the hidden dimensionality to 64, with layer normalization. Additional mesh input includes 3 channels for nodes and 4 for edges. The model’s processor features 12 layers, and a hidden layer dimension of 64. Table 2 provides results between all baselines and the FireCastNet model for timeseries length 24 for different time horizons. The TeleVit14 model is also included in the comparison. TeleViT is a teleconnection-driven vision transformer for seasonal wildfire forecasting, trained also on the SeasFire dataset. It should be noted that it does not work directly with timeseries data. Instead, it extracts temporal context from Oceanic Indices (OCIs). In all cases, different models were trained for each prediction time horizon. Regardless of the range of forecasts, the FireCastNet is clearly more potent in per-pixel classification.Fig. 4Performance of FireCastNet model at a global scale with timeseries length 24 for different time overlap windows. Performance of FireCastNet model at a global scale with timeseries length 24 for different time overlap windows. Timeseries length In this section, we investigate the impact of training global models using input time series of varying lengths, both with and without spatially accumulated information. This analysis aims to explore how the temporal context provided to the models affects their predictive performance across different forecasting horizons. Specifically, we evaluate models trained with three distinct time-series lengths, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ts \in \{6, 12, 24\}$$\end{document}, and analyze their performance in terms of the Area Under the Precision-Recall Curve (AUPRC). For each combination of time-series length and forecasting horizon \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h \in \{1, 2, 4, 8, 16, 24\}$$\end{document}, we train a separate model to assess the influence of these parameters independently. Figure 3 summarizes the performance of the models under these configurations, offering insights into how time-series length affects prediction accuracy. Among the various models studied, the GRU model demonstrates a clear positive correlation between time-series length and performance. The results indicate that longer input sequences provide richer temporal context, which enhances the GRU’s ability to capture complex temporal dependencies and improves predictive accuracy consistently across all forecasting horizons. For the Conv-LSTM model, a similar trend is observed, albeit with diminishing returns as the time-series length increases from 12 to 24. While both 12 and 24-length time series yield competitive results, the latter shows slight advantages, particularly for longer-term forecasts. On the other hand, the Conv-GRU model exhibits an unexpected behavior, where mid-range time-series lengths (12) outperform both shorter (6) and longer (24) inputs. This suggests that the Conv-GRU may balance the trade-off between overfitting to longer sequences and underutilizing shorter ones. The U-TAE model shows competitive performance with both 12 and 24-length time series, although the gap between these configurations widens for longer forecasting horizons. This behavior implies that while the U-TAE model benefits from longer input sequences for more extended predictions, it does not fully exploit the additional temporal information as effectively. FireCastNet demonstrates stability and robustness compared to the other architectures. Regardless of the forecasting horizon, the model maintains relatively consistent performance, particularly when trained with the longest time-series length of 24. This suggests that FireCastNet effectively leverages the full temporal context. Time overlap In this section, we conduct an ablation study to investigate the effect of varying sample overlaps between consecutive time series on model accuracy, as measured by the Area Under the Precision-Recall Curve (AUPRC). The goal is to understand the role of overlapping temporal context in improving model performance and its implications for training efficiency and predictive accuracy. In the original setup, for a time series of length t, each sample overlaps with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t-1$$\end{document} identical points from the subsequent sample. This ensures that consecutive input sequences share nearly the entire temporal context, differing by only a single time step. To systematically evaluate the effect of overlap, we consider scenarios where the overlap is reduced. Specifically, for time series with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t=24$$\end{document}, we analyze configurations where consecutive samples share 3, 6, 12, 18, and including the original full-overlap configuration of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t-1=23$$\end{document} samples The results, illustrated in Fig. 4, reveal a clear proportional relationship between the degree of overlap and the resulting AUPRC scores. Higher overlaps consistently yield better AUPRC values, underscoring the importance of preserving shared temporal information between samples. Using the highest possible overlap (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t-1=23$$\end{document}) offers several key advantages such as enabling the model to capture subtle temporal patterns and correlations, particularly for dynamic and time-sensitive phenomena and leveraging information from adjacent time steps. While higher overlaps do result in a greater computational burden due to increased redundancy in input samples, the benefits in terms of predictive accuracy and model robustness outweigh this drawback in many practical scenarios. The results emphasize that preserving shared temporal context is a crucial factor for achieving superior accuracy, particularly in tasks requiring fine-grained temporal predictions. Variable importance Fig. 5Feature attributions for fire-pattern predictions with the global model across forecasting horizons (1, 2, 4, 8, 16, 24 8-days). Each column sums to 1.0, indicating the relative importance of each variable for the respective horizon. The results were obtained using Integrated Gradients36. Feature attributions for fire-pattern predictions with the global model across forecasting horizons (1, 2, 4, 8, 16, 24 8-days). Each column sums to 1.0, indicating the relative importance of each variable for the respective horizon. The results were obtained using Integrated Gradients36. We applied a straightforward Integrated Gradients36 attribution to gain initial intuition on which input variables most influence FireCastNet’s predictions. For the computation of the gradients, we used a zero baseline and performed 200 steps. The resulting feature attributions for the global model are presented in Fig. 5. These attributions indicate the relative importance of each variable in determining the final fire-pattern prediction across different forecasting horizons. Each column in Fig. 5 corresponds to predictions for a specific horizon and sums up to 1.0. We evaluated the importance of nine dynamic variables and two static variables (population density and Land Sea Mask) across six prediction horizons: 1, 2, 4, 8, 16, and 24 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 8-day periods. These variables were selected to reflect the major components influencing wildfire activity, including atmospheric conditions, land surface and vegetation properties, oceanic influence and anthropogenic impacts. As expected, the mean sea level pressure (mslp)—a key indicator of large-scale weather patterns—and the Normalized Difference Vegetation Index (ndvi)—a proxy for vegetation health and fuel availability—emerge as dominant predictors across nearly all horizons. A consistently important secondary variable is the surface solar radiation downwards (ssrd), which quantifies the amount of sunlight reaching the ground. This variable modulates fire susceptibility through its control on surface energy balance, influencing fuel temperature, moisture loss via drying and evapotranspiration, and the timing and intensity of seasonal flammability patterns, with particularly strong effects in regions characterized by persistent high radiation and low moisture availability (e.g., savannas, Mediterranean ecosystems). Below, we summarize how variable importance evolves with forecasting lead time:Short-Term Forecasts (h = 1, 2, 4 8-day periods): The highest attributions are consistently found in ndvi and mslp. Land surface temperature at day (lst_day) also shows significant influence.Mid-to-Long-Term Forecasts (h = 8, 16 8-day periods): mslp strongly dominates these horizons, reaching its peak importance at h = 16 (0.240). ndvi and ssrd remain key, while variables like Volumetric soil water level 1 (swvl1) show their lowest importance.Longest-Term Forecasts (h = 24 8-day periods): While mslp remains the single most important variable (0.197), the attribution for ndvi generally decreases. The influence of lst_day and Total precipitation (tp) increases significantly compared to mid-range forecasts.Variables with consistently low attribution across all horizons include Sea surface temperature (sst), Population density (pop_dens), and the Land sea mask (lsm). This pattern suggests that while shorter-term predictions rely heavily on immediate environmental conditions (reflected in high initial NDVI and LST contributions), longer-term seasonal forecasts increasingly depend on broader atmospheric processes captured by mslp, and the combined influence of long-term moisture availability (tp) and surface heat (lst_day). Short-Term Forecasts (h = 1, 2, 4 8-day periods): The highest attributions are consistently found in ndvi and mslp. Land surface temperature at day (lst_day) also shows significant influence. Short-Term Forecasts (h = 1, 2, 4 8-day periods): The highest attributions are consistently found in ndvi and mslp. Land surface temperature at day (lst_day) also shows significant influence. Mid-to-Long-Term Forecasts (h = 8, 16 8-day periods): mslp strongly dominates these horizons, reaching its peak importance at h = 16 (0.240). ndvi and ssrd remain key, while variables like Volumetric soil water level 1 (swvl1) show their lowest importance. Mid-to-Long-Term Forecasts (h = 8, 16 8-day periods): mslp strongly dominates these horizons, reaching its peak importance at h = 16 (0.240). ndvi and ssrd remain key, while variables like Volumetric soil water level 1 (swvl1) show their lowest importance. Longest-Term Forecasts (h = 24 8-day periods): While mslp remains the single most important variable (0.197), the attribution for ndvi generally decreases. The influence of lst_day and Total precipitation (tp) increases significantly compared to mid-range forecasts. Longest-Term Forecasts (h = 24 8-day periods): While mslp remains the single most important variable (0.197), the attribution for ndvi generally decreases. The influence of lst_day and Total precipitation (tp) increases significantly compared to mid-range forecasts. Finally, it is important to note that explainability at the variable level remains challenging in multivariate Earth system forecasting. Attribution scores may not fully reflect a variable’s true physical influence when it is strongly correlated with others, since the contribution of a low-attribution predictor can be implicitly captured by another with higher attribution. For example, Volumetric soil water level 1 (swvl1), which represents the moisture content in the top soil layer and is closely linked to fuel moisture availability, shows consistently low attribution across horizons (Fig. 5). Rather than indicating low relevance, this likely reflects strong dependence on broader moisture and energy balance controls represented by variables such as precipitation and mean sea level pressure, rather than a lack of physical relevance to fire patterns. Global model per GFED regions Table 3Global model performance per GFED region for different prediction horizons.GFED regionFraction of burned areasHorizon12481624Boreal North America (BONA)0.924%0.040.030.030.030.020.02Temperate North America (TENA)1.986%0.260.270.260.260.250.26Central America (CEAM)2.137%0.550.540.530.530.550.55Northern Hemisphere South America (NHSA)2.673%0.670.650.640.630.610.61Southern Hemisphere South America (SHSA)15.619%0.440.430.430.420.410.41Europe (EURO)0.857%0.200.200.190.180.170.19Middle East (MIDE)1.029%0.290.300.290.300.290.26Northern Hemisphere Africa (NHAF)20.749%0.740.730.720.710.710.72Southern Hemisphere Africa (SHAF)29.988%0.850.840.840.830.830.84Boreal Asia (BOAS)4.072%0.140.130.130.130.130.14Central Asia (CEAS)8.264%0.270.260.270.260.250.24Southeast Asia (SEAS)5.764%0.640.630.620.610.600.61Equatorial Asia (EQAS)1.089%0.490.500.480.440.400.40Australia and New Zealand (AUST)4.849%0.310.310.300.300.290.32The fraction of burned areas represents the percentage of wildfires in a specific region relative to the total number of fires globally. Table 3Global model performance per GFED region for different prediction horizons.GFED regionFraction of burned areasHorizon12481624Boreal North America (BONA)0.924%0.040.030.030.030.020.02Temperate North America (TENA)1.986%0.260.270.260.260.250.26Central America (CEAM)2.137%0.550.540.530.530.550.55Northern Hemisphere South America (NHSA)2.673%0.670.650.640.630.610.61Southern Hemisphere South America (SHSA)15.619%0.440.430.430.420.410.41Europe (EURO)0.857%0.200.200.190.180.170.19Middle East (MIDE)1.029%0.290.300.290.300.290.26Northern Hemisphere Africa (NHAF)20.749%0.740.730.720.710.710.72Southern Hemisphere Africa (SHAF)29.988%0.850.840.840.830.830.84Boreal Asia (BOAS)4.072%0.140.130.130.130.130.14Central Asia (CEAS)8.264%0.270.260.270.260.250.24Southeast Asia (SEAS)5.764%0.640.630.620.610.600.61Equatorial Asia (EQAS)1.089%0.490.500.480.440.400.40Australia and New Zealand (AUST)4.849%0.310.310.300.300.290.32The fraction of burned areas represents the percentage of wildfires in a specific region relative to the total number of fires globally. Global model performance per GFED region for different prediction horizons. The fraction of burned areas represents the percentage of wildfires in a specific region relative to the total number of fires globally. Fig. 6Global model AUPRC performance across horizons per GFED region. Dashed lines represent the any-fire (red) and majority (green) baselines. Global model AUPRC performance across horizons per GFED region. Dashed lines represent the any-fire (red) and majority (green) baselines. In this section, we present predictions derived using regional masks from the Global Fire Emissions Database (GFED), as provided by the SeasFire Datacube16. These regional masks allow for a more granular analysis of fire activity by focusing on specific geographic areas. To produce region-specific predictions, we compute the Area Under the Precision-Recall Curve (AUPRC) for each region by applying the GFED masks to both the model’s output and the ground truth burned area data. The motivation for evaluating the model’s performance across different GFED regions stems from the fact that the performance of the global model, as can be seen in Table 2, does not always drop significantly as the prediction horizon increases. Thus, we hypothesize that the model’s performance may vary across different regions depending on the number of samples in our dataset. This is particularly relevant for regions with distinct fire regimes, where the model’s ability to generalize may be influenced by the availability of training data and the specific characteristics of each region’s fire patterns. Notable region is Africa, which has a significant number of wildfires. Table 3 and Fig. 6 summarize the AUPRC results across all GFED regions and for multiple time horizons, providing a comprehensive overview of the model’s performance in diverse settings. In the same table we have also added the fractions of burned areas per GFED region. These represent the percentages of wildfires in a specific region relative to the total number of fires globally, calculated from 2002-01-01 to 2020-01-01. Notably, FireCastNet demonstrates superior accuracy in burned area forecasts, particularly in Southeast Asia, Northern Hemisphere South America, and Africa, highlighting its effectiveness in regions with distinct fire regimes and challenges. On the other hand, some regions such as Northern Hemisphere Africa (NHAF) and Southern Hemisphere Africa (SHAF) exhibit different behavior than regions such as Northern Hemisphere South America (NHSA) and Southern Hemisphere South America (SHSA). NHAF and SHAF show a drop in performance as the prediction horizon increases, while NHSA and SHSA maintain relatively stable performance across horizons. This suggests that in the case of Africa the model seems to learn some advanced seasonality pattern instead of learning how to predict the burned areas. This is further supported by the fact that the model’s performance in Africa is significantly better than in other regions, such as Southeast Asia (SEAS) and Equatorial Asia (EQAS), where the model’s performance drops significantly as the prediction horizon increases. This discussion motivates “Local area modeling” section. Local area modeling The Local Area Modeling (LAM) approach4,5 extends the uniform global strategy by implementing region-specific training. Building upon the hierarchical multi-resolution mesh construction, where each target GFED region receives fine-grained mesh granularity while surrounding areas transition to progressively coarser levels, we train dedicated FireCastNet models for individual regions of interest. This localized training strategy offers several key advantages over the global model approach. First, by constraining the training domain to a specific region, the model can better capture the unique fire-climate relationships and local environmental drivers that characterize different biomes and geographical areas. Second, the computational focus on smaller geographical domains allows for more efficient use of the model’s capacity. Third, the LAM approach enables enhanced spatial resolution within the region of interest while maintaining computational tractability. The adaptive mesh strategy concentrates the finest mesh granularity within the target region, providing higher spatial fidelity for local fire prediction while the coarser surrounding mesh levels capture the necessary global teleconnections and boundary conditions.Fig. 7Example of mesh using the Local Area Modeling (LAM) approach for the SHSA GFED Region (Southern Hemisphere South America). The left image shows the flat mesh representation around the target region. While not shown the grid expands to the entire globe as can be seen in the right image, which shows the spherical mesh representation. The target region is represented at the sixth level of detail, while the surrounding areas are represented at the third level of detail. The mesh is adapted to focus on the target region, allowing for higher resolution predictions in that area while maintaining a coarser resolution in the surrounding regions. The transition from finer to coarser mesh levels is gradual, using buffer regions of size 400–800 km, ensuring a smooth transition and preserving the spatial coherence of the model’s predictions. Figure generated using the Matplotlib30 library (https://zenodo.org/records/17595503). Example of mesh using the Local Area Modeling (LAM) approach for the SHSA GFED Region (Southern Hemisphere South America). The left image shows the flat mesh representation around the target region. While not shown the grid expands to the entire globe as can be seen in the right image, which shows the spherical mesh representation. The target region is represented at the sixth level of detail, while the surrounding areas are represented at the third level of detail. The mesh is adapted to focus on the target region, allowing for higher resolution predictions in that area while maintaining a coarser resolution in the surrounding regions. The transition from finer to coarser mesh levels is gradual, using buffer regions of size 400–800 km, ensuring a smooth transition and preserving the spatial coherence of the model’s predictions. Figure generated using the Matplotlib30 library (https://zenodo.org/records/17595503). Figure 7 illustrates the adaptive mesh strategy employed in the LAM approach, highlighting the varying levels of granularity across different regions. The figure demonstrates how the mesh is refined within the target region while coarser resolutions are used in surrounding areas, effectively balancing detail and computational efficiency. We reach the sixth level of detail in the target region, while the surrounding areas are represented at the third level of detail (levels here correspond to the different mesh resolutions in the original GraphCast mesh construction). The change from finer to coarse mesh levels is gradual, ensuring a smooth transition and preserving the spatial coherence of the model’s predictions.Table 4Local area model performance per GFED region for different prediction horizons.GFED RegionFraction of burned areasHorizon12481624Boreal North America (BONA)0.924%0.040.040.030.020.030.03Temperate North America (TENA)1.986%0.280.270.270.280.290.29Central America (CEAM)2.137%0.400.390.370.380.390.38Northern Hemisphere South America (NHSA)2.673%0.560.550.550.540.540.54Southern Hemisphere South America (SHSA)15.619%0.470.470.470.460.450.44Europe (EURO)0.857%0.240.240.240.210.240.20Middle East (MIDE)1.029%0.310.310.290.330.290.28Northern Hemisphere Africa (NHAF)20.749%0.780.770.750.740.750.75Southern Hemisphere Africa (SHAF)29.988%0.860.860.860.850.850.85Boreal Asia (BOAS)4.072%0.090.100.070.060.060.07Central Asia (CEAS)8.264%0.310.310.310.300.280.29Southeast Asia (SEAS)5.764%0.570.570.550.550.520.54Equatorial Asia (EQAS)1.089%0.500.490.480.460.470.48Australia and New Zealand (AUST)4.849%0.250.260.240.240.250.26A different model is trained for each region, using the same training data as the global model, but focusing on the region of interest. Table 4Local area model performance per GFED region for different prediction horizons.GFED RegionFraction of burned areasHorizon12481624Boreal North America (BONA)0.924%0.040.040.030.020.030.03Temperate North America (TENA)1.986%0.280.270.270.280.290.29Central America (CEAM)2.137%0.400.390.370.380.390.38Northern Hemisphere South America (NHSA)2.673%0.560.550.550.540.540.54Southern Hemisphere South America (SHSA)15.619%0.470.470.470.460.450.44Europe (EURO)0.857%0.240.240.240.210.240.20Middle East (MIDE)1.029%0.310.310.290.330.290.28Northern Hemisphere Africa (NHAF)20.749%0.780.770.750.740.750.75Southern Hemisphere Africa (SHAF)29.988%0.860.860.860.850.850.85Boreal Asia (BOAS)4.072%0.090.100.070.060.060.07Central Asia (CEAS)8.264%0.310.310.310.300.280.29Southeast Asia (SEAS)5.764%0.570.570.550.550.520.54Equatorial Asia (EQAS)1.089%0.500.490.480.460.470.48Australia and New Zealand (AUST)4.849%0.250.260.240.240.250.26A different model is trained for each region, using the same training data as the global model, but focusing on the region of interest. Local area model performance per GFED region for different prediction horizons. A different model is trained for each region, using the same training data as the global model, but focusing on the region of interest. Training these localized models involves adapting the FireCastNet loss function to compute the loss only on the region of interest. Table 4 presents the results. Again each table entry corresponds to a separate model trained on the region of interest for that particular prediction horizon, using the same training data as the global model. The results can be compared with the global model results presented in Table 3, indicating a varied impact on performance across different GFED regions. Several regions exhibit a notable improvement in AUPRC scores with the LAM models. For instance, regions in Africa (NHAF and SHAF), Europe (EURO), the Middle East (MIDE), and parts of the Americas (TENA, SHSA) and Asia (EQAS, CEAS) show enhanced predictive performance. Conversely, for some regions, the global model outperforms the specialized LAM models. This is observed in regions such as Boreal North America (BONA), Central America (CEAM), Northern Hemisphere South America (NHSA), Boreal Asia (BOAS), Southeast Asia (SEAS), and Australia (AUST). In these cases, it is plausible that long-range teleconnections, which are better captured by the global model, play a more dominant role in driving fire activity than local factors. The global model’s ability to learn from a wider range of fire regimes and climatic conditions across the entire planet may provide a more robust and generalizable prediction for these specific areas. This highlights a key trade-off between local specialization and the comprehensive global context. Examining the relationship between LAM performance improvements and the fraction of burned areas in each region (which serves as a proxy for data availability), we find no clear correlation between the number of samples and whether LAM models outperform the global model. While some high-activity regions like Africa (NHAF, SHAF) benefit from LAM, other regions with substantial fire activity such as Southeast Asia (SEAS) and Australia (AUST) do not show improvement. Similarly, regions with limited fire activity like Europe (EURO) and Temperate North America (TENA) still benefit from the LAM approach. This lack of observable correlation suggests that the effectiveness of LAM models depends on factors beyond simple data availability, likely related to the underlying fire dynamics and their relationship to local versus global climate drivers in each region. Conclusion The findings of this study underscore the potential of machine learning models in advancing seasonal wildfire forecasting. Our architecture, FireCastNet, integrates 3D convolutional encoding with GraphCast, in order to capture the spatio-temporal context of wildfires, enabling predictive analysis across various forecasting horizons. Our architecture is clearly more beneficial than both baseline and contemporary methods for spatio-temporal predictions and burned area prediction via semantic segmentation. One significant insight derived from our investigation is the importance of temporal context. The results clearly show that longer input time-series enhance predictive robustness across forecasting horizons. This finding aligns with the intuition that wildfires are influenced by accumulated climatic, vegetative, and human-related factors, which can be utilized more effectively by adding a deeper temporal history. This outcome shows the benefits of architectures that can process both local and regional patterns, highlighting the interconnected nature of wildfire phenomena where adjacent regions’ conditions can influence fire behavior.