Optimization of Convolutional Neural Networks for Hedgerow Object Detection in Very High-Resolution Satellite Images

Kurzfassung

Hedgerows are one of the few remaining natural landscape features within agricultural areas. They provide important ecosystem services, living space for numerous species, and mitigate habitat fragmentation by acting as corridors between habitat patches. Environmental agencies in many European countries are thus interesting in monitoring and protecting hedgerows. In order to facilitate the success of such monitoring programs, cost-effective and accurate mapping of hedgerows is required. Here, the use of neural networks is examined in order to determine their feasibility for hedgerow mapping. Two state of the art neural networks were chosen (Mask R-CNN and DeepLab v3+). Both networks were able to successfully detect hedgerows, with DeepLab v3+ outperforming Mask R-CNN. Images from two different seasons (October and May) were tested as inputs. Finding suggest that using all available images, regardless of season, is preferred. Data augmentations were found to greatly increase performance for both networks, with DeepLab v3+ achieving up to 81% recall and 69% precision. However, caution should be taken when applying augmentations which modify spectral values of pixels. A pre-trained network was compared to a network with randomly initialized weights using DeepLab v3+. Here, the pre-trained network was found to produce superior recall and precision. However, the mask boundaries were far more precise using the randomly initialized network. The main limitation to successfully using randomly initialized weights was the amount of data available. Improvements can be made to Mask R-CNN by using rotatable anchors, as the network mainly struggled to detect diagonal hedgerows. Both neural networks were able to perform detections across a large spatial scale, and compared to previous object-based hedgerow mapping techniques, were relatively simple to use. Thus, this study supports the use of neural networks use in regional monitoring strategies for hedgerows