Advanced Global Wildfire Activity Modeling with Hierarchical Graph ODE
Authors: Fan Xu, Wei Gong, Hao Wu, Lilan Peng, Nan Wang, Qingsong Wen, Xian Wu, Kun Wang, Xibin Zhao
Organizations: University of Science and Technology of China, Hefei, Anhui, China · Tsinghua University, Beijing, China · Southwest Jiaotong University, Chengdu, Sichuan, China · Beijing Jiaotong University, Beijing, China · Squirrel AI, Seattle, Washington, USA · Tencent, Beijing, China · Nanyang Technological University, Singapore
Wildfires, as an integral component of the Earth system, are governed by a complex interplay of atmospheric, oceanic, and terrestrial processes spanning a vast range of spatiotemporal scales. Modeling their global activity on large timescales is therefore a critical yet challenging task. While deep learning has recently achieved significant breakthroughs in global weather forecasting, its potential for global wildfire behavior prediction remains underexplored. In this work, we reframe this problem and introduce the Hierarchical Graph ODE (HiGO), a novel framework designed to learn the multi-scale, continuous-time dynamics of wildfires. Specifically, we represent the Earth system as a multi-level graph hierarchy and propose an adaptive filtering message passing mechanism for both intra- and inter-level information flow, enabling more effective feature extraction and fusion. Furthermore, we incorporate GNN-parameterized Neural ODE modules at multiple levels to explicitly learn the continuous dynamics inherent to each scale. Through extensive experiments on the SeasFire Cube dataset, we demonstrate that HiGO significantly outperforms state-of-the-art baselines on long-range wildfire forecasting. Moreover, its continuous-time predictions exhibit strong observational consistency, highlighting its potential for real-world applications.
Figures & tables
Figure 1. The left depicts a local wildfire event, influenced by immediate conditions like airstreams. The right illustrates the broader Earth system context, framing wildfires as a phenomenon governed by a complex interaction between the Atmosphere, Ocean, Land and Human Activity.
Figure 2. Architectural overview of the HiGO framework. The framework integrates multi-modal driver and climate data to initialize a hierarchical graph representation. The core of the model consists of a multi-scale graph pyramid, where HGODE blocks learn continuous-time dynamics and an adaptive message-passing scheme captures complex spatial dependencies.
8 days
16 days
24 days
32 days
40 days
48 days
Category
Model
M-F1
AUPRC
M-F1
AUPRC
M-F1
AUPRC
M-F1
AUPRC
M-F1
AUPRC
M-F1
AUPRC
Point-wise
MLP [9]
0.451
0.473
0.408
0.382
0.361
0.348
0.320
0.277
0.281
0.248
0.265
0.223
XGBoost [2]
0.439
0.485
0.429
0.432
0.355
0.377
0.312
0.320
0.303
0.291
0.288
0.266
Vision-based
U-Net [8]
0.495
0.551
0.449
0.498
0.407
0.452
0.369
0.411
0.334
0.375
0.322
0.343
ViT [3]
0.543
0.623
0.478
0.587
0.465
0.518
0.417
0.458
0.364
0.441
0.337
0.406
SwinT [7]
0.561
0.627
0.527
0.604
0.486
0.535
0.438
0.493
0.382
0.475
0.356
0.430
Table 1. Long-term forecasting performance comparison across multiple lead times. We report Macro F1-Score (M-F1) and Area Under the Precision-Recall Curve (AUPRC). The best-performing model in each column is highlighted in bold , and the second-best is underlined . HiGO consistently demonstrates superior performance, with the advantage becoming more pronounced at longer horizons.
Figure 3. Visualization evaluation of HiGO’s long-term wildfire activity forecasting performance. The panel provides a visual comparison between HiGO’s multi-step predictions (top row) and the ground truth (bottom row) for up to a 48-day horizon, demonstrating high qualitative fidelity.
8-day
24-day
Model
M-F1
AUPRC
M-F1
AUPRC
GraphCast (Interp)
0.484
0.492
0.459
0.512
HiGO (Interp)
0.508
0.524
0.477
0.554
NDCN (Conti)
0.493
0.537
0.486
0.561
HiGO (Conti)
0.551
0.623
0.507
0.597
Table 2. Quantitative evaluation of continuous-time flexibility. Models are trained on a 16-day horizon. Performance is then measured on an intermediate, unseen time-step (8-day & 24-day)). Best results are in bold .
Figure 4. Impact of Forcing Variables on 24-day Forecast Performance. The bar chart shows the AUPRC score for the full model (left) and five variants, each with one group of input variables removed.
16-day
32-day
Filter Type
Mechanism
M-F1
AUPRC
M-F1
AUPRC
Low-pass
GCN
0.521
0.585
0.459
0.526
High-pass
FAGCN
0.530
0.598
0.468
0.534
Band-pass
BWGNN
0.539
0.607
0.475
0.542
Adaptive (Ours)
HiGO
0.548
0.619
0.484
0.563
Table 3. Ablation study of the message-passing mechanism. Our adaptive approach is benchmarked against GNNs with fixed low-pass, high-pass, and band-pass filters on 16-day and 32-day forecast horizons. Best results are in bold.
Figure 5. Impact of the number of hierarchical graph levels ( L ) on forecasting performance.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Full name
Pre-processing
Local/Global Variables
Mean sea level pressure
Total precipitation
Log-transformed
Vapour Pressure Deficit
Sea Surface Temperature
Mean Temperature at 2 meters
Appendix
Table 4. Input and target variables used from the SeasFire cube for all settings. The same variables are used for both local and global views.
Wildfire forecasting predicts where fires will start, how they will spread, and their likely intensity or impacts using environmental and fire observations.
Long-horizon wildfire risk forecasting requires generating probabilistic spatial fields under sparse event supervision while maintaining computational efficiency across multiple prediction horizons. Extending diffusion models to multi-step forecasting typically repeats the denoising process independently for each horizon, leading to redundant computation. We introduce N-Tree Diffusion (NT-Diffusion), a hierarchical diffusion model designed for long-horizon wildfire risk forecasting. Fire occurrences are represented as continuous Fire Risk Maps (FRMs), which provide a smoothed spatial risk field suitable for probabilistic modeling. Instead of running separate diffusion trajectories for each predicted timestamp, NT-Diffusion shares early denoising stages and branches at later levels, allowing horizon-specific refinement while reducing redundant sampling. We evaluate the proposed framework on a newly collected real-world wildfire dataset constructed for long-horizon probabilistic prediction. Results indicate that NT-Diffusion achieves consistent accuracy improvements and reduced inference cost compared to baseline forecasting approaches.
Yucheng Xing, Xin Wang
Department of Electrical and Computer Engineering, Stony Brook University, Stony Brook, NY 11794, USA
Background: Wildfires in Canada present increasing threats to ecosystems, communities, and infrastructure, demanding accurate forecasting tools to aid mitigation efforts. Existing models often lack scalability or fail to capture temporal dynamics effectively. Aims: This study aims to develop a deep learning framework tailored to Canadian wildfire spread prediction that captures spatio-temporal patterns in environmental data. Methods: We propose a U-Net architecture integrating a Video Swin Transformer encoder with a convolutional decoder to model three-day sequences of meteorological and environmental variables. Data are exclusively sourced from public repositories via Google Earth Engine, ensuring transparency and scalability. The model is trained and tested on a curated dataset of major Canadian wildfire events from 2014 to 2023. Key results: Our approach achieves strong predictive performance by effectively leveraging spatio-temporal attention to forecast next-day fire incidence maps. Conclusions: The model successfully captures complex wildfire dynamics unique to Canada's landscape and temporal variability. Implications: This framework paves the way for advanced spatio-temporal wildfire forecasting research and operational applications using publicly accessible datasets.
Maulik Srivastava, Esha Saha, Hao Wang
Faculty of Science, University of Alberta · Interdisciplinary Lab for Mathematical Ecology & Epidemiology (ILMEE), University of Alberta · Department of Mathematical and Statistical Sciences, University of Alberta
Benchmarks such as WildfireSpreadTS (WSTS) have enabled data-driven forecasting of wildfire spread from multimodal satellite time series, but they suffer from limitations: size restricts them to large fires that burn for weeks, and they are from similar environments. We introduce AFSeq, a multimodal dataset of 1,194 recent fire events (2023 to 2026) from 8 different countries across three regions, dominated by the short, small fires that existing benchmarks underrepresent. Our dataset is the largest and most diverse to date for time-series-based wildfire spread prediction. We also introduce UTAE-AR, an extension of a classical U-Net with Temporal Auto Encoder, enhanced with recurrent modules before the decoder. This allows the model to not only predict the next day but the whole sequence in an autoregressive way. Moreover, our model can feed its own predicted fire mask back one day at a time, producing the day-by-day forecasts that single-step models cannot. Our results show that UTAE-AR surpasses state-of-the-art on both WSTS, WSTS+ and AFSeq for next-day accuracy while having 21.5% of their parameters, and being able to predict confidently at a longer horizon.
Jose Guillen, Renzo Zanca, Valentin Barriere
CENIA, Macul, Chile · Universidad de Chile, DCC, Santiago, Chile