Wide-Area Land Cover Mapping with Sentinel-1 Imagery Using Deep Learning Semantic Segmentation Models

Land cover (LC) mapping is essential for monitoring the environment and understanding the effects of human activities on it. Recent studies demonstrated successful applications of specific deep learning models to small-scale LC mapping tasks (e.g., wetland mapping). However, it is not readily clear which of the existing state-of-the-art models for natural images are the best candidates to be taken for the particular remote sensing task and data. In this article, we answer that question for mapping the fundamental LC classes using the satellite imaging radar data. We took ESA Sentinel-1 C-band SAR images acquired during the whole summer season of 2018 in Finland, which are representative of the land cover in the country. CORINE LC map was used as a reference, and the models were trained to distinguish between the five major CORINE-based classes. We selected seven among the state-of-the-art semantic segmentation models so that they cover a diverse set of approaches: U-Net, DeepLabV3+, PSPNet, BiSeNet, SegNet, FC-DenseNet, and FRRN-B, and further fine-tuned them. Upon evaluation and benchmarking, all the models demonstrated solid performance with overall accuracy between 87.9% and 93.1%, with good to a very good agreement (Kappa statistic between 0.75 and 0.86). The two best models were fully convolutional DenseNets (FC-DenseNet) and SegNet (encoder-decoder-skip), with the latter having a much shorter inference time. Overall, our results indicate that the semantic segmentation models are suitable for efficient wide-area mapping using satellite SAR imagery and provide baseline accuracy against which the newly proposed models should be evaluated.