Signal, Image and Video Processing Обработка сигналов

Predicting the ROS of wildland fire using a deep neural network with a hybrid architecture

Lin, Chuanying, Li, Xingdong, Geng, Yingda, Zhao, Hongyang, Li, Dandan
Springer Nature · 2026-11-16 · DOI 10.1007/s11760-026-05879-1

Аннотация

The normal rate of spread (ROS) describes the outward motion of a wildfire front. We constructed a pixel-level ROS dataset for the Maoer Mountain region from 300 completed, 48-h FARSITE surface-fire simulations, crossing 100 ignition points with three historical weather windows from the fifth-generation European Centre for Medium-Range Weather Forecasts reanalysis (ERA5). A hybrid neural network combines four spatial landscape channels with five current scalar features to predict the mean normal ROS over 1 h. Ignition locations, rather than pixels, define the training, validation and test split. With a fixed 36 $$\times $$ × 36-pixel input neighborhood, the hybrid model achieved a mean absolute error (MAE) of 0.1035 m/min and a root-mean-square error (RMSE) of 0.3386 m/min on 2,716,385 accepted test samples from 58 simulated fires at held-out ignition locations. Compared with an information-matched multilayer perceptron (MLP) trained under the same protocol, the hybrid reduced test MAE by 43.7% and RMSE by 31.6% in a single-seed comparison. Errors differed markedly by spread rate: MAE was 0.0685 m/min below 1 m/min but 0.6957 m/min at or above that threshold. Across nine paired simulated scenarios, lowering assigned fuel moisture increased 48-h burned area by a mean of 27.5%, while raising it decreased area by 22.4%. These results characterize a surrogate for simulated surface-fire spread under the supplied landscape and weather assumptions.

Читать у издателя →