A Method of Training Neural Networks to Extract Wind-formed Sand Ripples

Abstract

Sand ripples are the smallest landforms in arid and semi-arid areas, and are extremely important for the study of wind-induced sand movement. They can be more conveniently measured using neural network digital image processing technology. This study extracted sand ripples using a combination of DenseNet and photos of aeolian sand ripples ridge lines. The study area was located at the junction between northwest Zhongwei and the southeastern edge of the Tengger Desert in the Ningxia Hui Autonomous Region. A convolutional neural network was trained using the ridge line image of aeolian sand ripples. After several iterations, a clear image was obtained. This paper provides a training model that can automatically monitor each frame in an image and provides a feasible scheme for the automatic monitoring of the formation of wind ripple ridges. The study has a certain reference value for the future construction of digital desert information.