Deep learning based windthrow detection for winter storms

Kurzfassung

Forests are an integral part of the natural environment providing social, economic, and environmental benefits. Though storms are an important part of natural forest dynamics, large magnitude storms can lead to uprooting of trees, known as windthrows. Post-storm management relies on proper and fast detection of windthrows. In this work, we study the detection of windthrows due to storm David in the coniferous forests of southern Lower-Saxony, Germany as an image segmentation problem. Two deep learning methods, previously researched U-Nets and current state-of-the-art DeepLabv3+ are compared. Often storm damaged forests are surveyed many months later under good weather conditions, however, we study a winter storm surveyed in winter conditions 19 days after the storm. Moreover, we generate a detailed prediction map by segmenting the input scenery into four classes, namely, no forest, forest with no windthrows, forests with windthrows, and cleared areas. The data consists of four spectral channels and we study different 3-channel combinations and input image tile sizes to obtain the best configuration for windthrow detection. DeepLabv3+ is found to outperform U-Net with a prediction accuracy of 86.27% for windthrows, with best accuracy of 95.03% across all classes, and a class IoU of 0.7440 compared to a prediction accuracy of 78.66% and class IoU of 0.6892 for U-Nets. Deeplabv3+ was able to process 2048 × 2048 mosaics with input image tile size of 512 × 512 in nearly 889ms. Thus, a fast and well performing windthrow detection model based on DeepLabv3+ is developed