Int J Performability Eng ›› 2018, Vol. 14 ›› Issue (11): 2820-2830.

### Remote Sensing Identification of Black Cotton Soil based on Deep Belief Network

Lingling Wang, Wenyin Gong*, and Xiang Li

1. School of Computer Science and Technology, China University of Geoscience, Wuhan, 430000, China
• Submitted on  ;
• Contact: * E-mail address: winniay@163.com
• About author:Lingling Wang is a doctoral student in the School of Computer Science and Technology at China University of Geoscience, Wuhan, Hubei. She received her Master’s degree from China University of Geoscience in 2013. Her current research interests include computer networks and security and machine learning.Wenyin Gong is a professor and doctoral supervisor. His main research interests include intelligent computing and its applications. He is currently the deputy secretary general of Hubei Computer Society and editor of the international SCI journal Memetic Computing. He has presided over two projects of the National Natural Science Foundation and one of the New Teachers’ Funds for Doctoral Programs under the Ministry of Education. Xiang Li is an associate professor and doctor of geoscience information engineering at China University of Geoscience.

Abstract: As a type of expansive soil, black cotton soil swells when absorbing water and shrinks when dehydrated, and the cycle of swelling-shrinking movements can readily occur repeatedly. These characteristics result in serious consequences both to land surfaces and to surface buildings such as ground fracturing, building settling, and road buckling and cracking, having extreme adverse effects on the quality and safety of road transportation. With Kitui, Kenya as the research area and a GF-1 remote sensing image as the vector, this study focuses on in-depth exploration of the application of a deep belief network to identify and classify black cotton soil based on the characteristics of the local black cotton soil in the remote sensing image. The results indicate that given the sample database available to this study, when the network depth was 3, the number of nodes in each hidden layer was 60, the learning rate was 0.01, the number of iterations was 20, and the number of samples was 2,000,000. The best classification result could be achieved with a precision of about 90% per the evaluation criteria proposed in this study, indicating a significant advantage of the deep belief network in remote sensing identification of black cotton soil.