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