SAR-DEM-Optical Mountainous dataset

SAR-optical remote sensing couples are widely exploited for their complementarity for land-cover and crops classifications, image registration, change detections and early warning systems. Nevertheless, most of these applications are performed on flat areas and cannot be generalized to mountainous regions. Indeed, steep slopes are disturbing the range sampling which causes strong distortions in radar acquisitions - namely, foreshortening, shadows and layovers. In order to strengthen the abilities of deep neural networks to perform in such extreme conditions, we propose a new dataset composed of 53.475 triplets of SAR, DEM and optical images acquired over the French Alps. SAR images from Sentinel-1 and optical images from Sentinel-2 have been standardly processed and registered with respect to the provided SRTM 1sec HGT DEM. The latter is also part of the dataset to give insights on the locations of strong slopes and therefore of distorted areas to the neural networks. Finally, foreshortening and shadow masks are released in order to evaluate networks performance on these areas. The dataset is provided to be used for unsupervised multimodal approaches or for transfer learning purposes to make standard neural networks generalizable and exploitable despite strong radar distortions. 

合成孔径雷达-光学遥感配对数据（SAR-optical remote sensing couples）因其互补性而被广泛应用于土地覆盖与作物分类、图像配准、变化检测及预警系统等领域。然而，这些应用大多局限于平坦区域，无法推广至山区。实际上，陡坡会干扰距离采样，导致雷达数据获取中出现严重畸变——即透视收缩（foreshortening）、阴影（shadows）和叠掩（layovers）。为增强深度神经网络（deep neural networks）在这类极端条件下的性能，我们提出了一个新数据集，该数据集包含53475组覆盖法国阿尔卑斯山区的合成孔径雷达（SAR）、数字高程模型（DEM）及光学图像三元组（triplets）。来自哨兵1号（Sentinel-1）的SAR图像和哨兵2号（Sentinel-2）的光学图像已按标准流程处理，并基于所提供的SRTM 1秒分辨率HGT数字高程模型（DEM）完成配准。后者（指DEM）亦包含于数据集中，旨在为神经网络提供陡坡位置及相应畸变区域的信息。最后，我们还提供了透视收缩和阴影掩膜（masks），用于评估网络在这些区域的性能。该数据集可用于无监督多模态方法（unsupervised multimodal approaches）或迁移学习（transfer learning），以提升标准神经网络在强雷达畸变条件下的泛化能力与可用性。