A multi-source remote sensing benchmark for event-level wildfire spread prediction in California using spatiotemporal deep learning

Satellite-based wildfire spread studies differ in data organization and evaluation procedures, making model performance difficult to compare. We constructed an event-level benchmark to compare model performance and examine how the models used the available inputs. The benchmark covers California during 2012 to 2021 and combines VIIRS 375 m active-fire observations, MODIS NDVI-derived vegetation information, and ERA5-Land wind fields. Each wildfire event was represented on a fixed 100 × 100 grid. Each sequence sample used ten historical daily steps to predict the next-day satellite-observed fire-state mask. Six spatiotemporal models (CNN-LSTM, Vision Transformer, ConvLSTM, Pairwise-ConvLSTM, Patchwise-ConvLSTM, and ST-GNN) were evaluated using common input definitions, preprocessing procedures, prediction targets, and metrics. Integrated Gradients was applied after training to quantify attributions from the five raster input channels to the next-day model outputs. Pairwise-ConvLSTM achieved the lowest overall RMSE of 0.0495 on the cumulative wildfire dataset. Among the five evaluated Patchwise-ConvLSTM configurations, the 7 × 7 setting achieved the highest observed F1-score (0.955), with precision (0.962) and recall (0.949). On the newly burned-pixel subset, the same configuration achieved precision (0.916), recall (0.888), and F1-score (0.902). The attribution analysis showed larger mean absolute values for vegetation-constrained unburned states (7.23 × 10−5) and prior cumulative fire history (5.13 × 10−5) than for newly detected fire pixels (1.81 × 10−5), the eastward wind component (7.01 × 10−6), and the northward wind component (8.97 × 10−6). The benchmark provides a common framework for assessing predictive performance and model input use in next-day satellite-observed wildfire spread prediction.