01. The Challenge
Wildfire risk models are frequently opaque and rarely disclose where they actually fail, making it hard to trust them for real decisions.
02. The Solution
Built an end-to-end pipeline fusing Sentinel-1/2, Landsat, and MODIS satellite imagery with ERA5 reanalysis weather through a CNN + attention encoder feeding a BiLSTM, producing interpretable 1/3/6-month wildfire risk classifications for two fire-active Alaska regions.
Overview
FireSat-AI is a GSoC-style MVP for the "Alaska Wildfire Prediction Using Satellite Imagery" proposal: a hybrid CNN-LSTM + attention model that fuses multi-source satellite imagery with weather reanalysis data to forecast wildfire risk at 1, 3, and 6 month horizons.
Architecture
Monthly satellite feature stacks pass through a ResNet-style CNN with squeeze-excite channel attention, producing a spatial embedding per month. These are concatenated with ERA5 weather features and fed through a bidirectional LSTM over a 24-month lookback, with additive temporal attention producing the final context vector for three multi-horizon classification heads.
Interpretability by Design
Both the channel attention and temporal attention are surfaced directly in the API and dashboard, alongside gradient x input saliency maps, so a predicted risk level always comes with an explanation of which regions and months drove it.
Honesty Notes
The shipped demo ships with a physically-motivated synthetic dataset (real Alaska climate normals, seasonal vegetation cycles) rather than live satellite pulls, and the evaluation report plainly states where the demo checkpoint does and doesn't beat a naive majority-class baseline — reported rather than hidden.