Classifying surface fuel types based on forest stand photographs and satellite time series using deep learning

With the increasing threat of wildfires globally, improving availability accurate, spatially explicit fuel type information is critical for fire behavior predictions that can support management decisions to mitigate hazards. Since mapping surface types using airborne or spaceborne sensors relies on ground truth data from laborious field assessments, here we propose a novel proximate sensing-based approach classifying in-forest RGB photographs convolutional neural networks (CNNs). We test different configurations deep learning models integrate forest stand and floor as well time series multispectral satellite Sentinel-2 long short-term memory (LSTM), compare their performance in understory litter Central European forests. also investigate how ensemble approaches based majority voting help improve classification results. found were classified with highest accuracy after cross-validation (0.78) combination horizontal photos photos. This was further improved by post-classification decision fusion model multiple considering model’s confidence its (0.85). Litter resulted lower overall (0.60), but both significantly results (0.72). our mostly limited naturally smooth transitions between defined classes co-occurrence photograph. study shows methods provide an efficient means assess GNSS-located stands basis generating validating finally risk maps. The necessary be readily collected managers citizen scientists.