Automated Early Detection of Wildfire Smoke Using Deep Learning with Combined Spatial-Temporal Information

Abstract:

Successful wildfire intervention is dependent on the early detection of wildfire smoke; mere minutes can drastically alter the extent of societal and ecological damage from wildfires. Automated early detection of wildfire smoke using deep learning models has shown promising results, but false positive rates remain high, particularly when the models are deployed to novel environments. To date, these models largely use Convolutional Neural Networks (CNNs) and other spatially-oriented computer vision techniques to detect smoke. In deep learning, temporal patterns can be learned via a Long Short-Term Memory network (LSTM). We propose incorporating both spatial and temporal information via a combined CNN-LSTM classification model. We theorize that the inclusion of temporal information may reduce the number of false positives and improve generalizability to new environments. The model is trained and tested on images of landscapes with and without smoke from the HPWREN tower network in southern California, part of the SAGE remote-sensing infrastructure. We use traditional CNN-based classifiers leveraged in past smoke detection literature as baselines to evaluate our model’s performance.