An Auto-Detection and classification algorithm for identification of sand dunes based on remote sensing images

Abstract

Dunes are among the most abundant aeolian landforms found on Earth and other planets. The automatic mapping of dunes over large-scale areas is crucial for predicting, monitoring, and managing lands threatened by sand encroachment. However, the spectral similarity between dune and inter-dune areas, as well as the diverse presentations of dunes of the same type, have made it challenging to automatically classify and map dunes. In this study, we propose a two-step 'detect-then-classify' framework that enables the automatic identification of heterogeneous dune types. The first step involves dune detection. We introduce a newly developed convolutional neural network called SandUnet, which is adapted from Attention U-Net. This network is designed to preserve uncompressed input signals ensuring that minor differences in the color and texture of the dunes are retained. The second step focuses on dune-type classification. To classify dunes, a fine-tuned MobileNet was constructed to integrate the wealth of knowledge ingrained in MobileNet's pre-trained layers with the adaptations tailored specifically for the sand dune images. Subsequently, each dune image is automatically classified into six different types using this framework. By applying this framework to all the bare sand areas of the Taklamakan (or Taklimakan) Desert using Landsat-8 imagery, the final classification of dune types of the desert was achieved. Testing on randomly selected 10% image patches of whole desert, the overall accuracy of the dune-type classification is 70%. This study presents an automatic and economical solution for identifying dunes in large-scale areas.