A spatio-temporal model for wildfire spread prediction in Thailand using multi-scale spatial features and cyclical temporal encoding

Information from the abstract

Accurate wildfire spread prediction is essential for effective mitigation in Thailand, where fires are frequent, small, and often arise from separate ignition sources. Existing research, largely trained on datasets from other regions, does not capture Thailand's distinct fire behavior, and most approaches rely solely on spatial models. We address this limitation by constructing a Thailand-specific wildfire dataset collected from 2024 to 2025, covering major forest regions across the country, and by developing a preprocessing strategy that removes newly ignited fire clusters using a density-based spatial clustering of applications with noise (DBSCAN) algorithm, ensuring that the model learns true propagation patterns. We apply the spatio-temporal model dubbed Simpler Yet Better Video Prediction (SimVP) for 1–3 day forecasting, enhanced with an Atrous Spatial Pyramid Pooling (ASPP) module to improve multi-scale spatial feature extraction and a cyclical temporal encoding (CTE) to incorporate seasonal cues. Experiments on our dataset show that our model achieves an IoU of 0.685 and an average F1-score of 0.718, outperforming all spatial and spatio-temporal baselines, with performance improvements most evident in 1–3-day-ahead forecasts. These results demonstrate the effectiveness of tailored preprocessing and multi-scale spatio-temporal modeling for wildfire forecasting in Thailand.

Why this record is monitored

This record has an Impact Signal of 72/100 based on recency, source, collaboration, and bibliographic signals. It prioritizes monitoring and is not a judgment of research quality.

Related topics: Fire effects on ecosystems · Fire Detection and Safety Systems · Landslides and related hazards