Automatic Sand Dune Detection and Classification Framework Using Remote Sensing Images

Abstract

<p>Dune field landscape patterns serve as an important signs of aeolian processes, such as wind conditions, sediment supply, and so on. A novel framework was proposed and evaluated for automatic dune detection and classification with remotely sensed images. The framework consists of two main steps: (1) The first step is to detect sand dunes from remote sensing images by SandUnet, which is firstly proposed in this paper. SandUnet, a Convolutional Neural Network (CNN), has a similar network structure with Attention U-net but modifies its attention gate module. In SandUnet, the input signals' information is not compressed as in the Attention U-net, therefore, the nuanced color and texture information of dunes are preserved. This paper demonstrated that SandUnet has better detection accuracy than other popular CNNs such as FCN, U-net, U-net++, and Attention U-net. (2) The second step is to compute the image similarity scores through MobileNet between each dune detection result image and the representative images of 6 different types of dunes. Then, each dune detection result image is classified into a dune type automatically. This paper applied the proposed framework to Taklimkan Desert in China. The average classification accuracy rate is around 80%, which proves the usefulness of this framework in automatic, no-cost, and accurate sand dune classification.</p>