AGA-SVR-based selection of feature subsets and optimization of parameter in regional soil salinization monitoring. Issue 12 (17th June 2020)

• Record Type: Journal Article
• Authors: Xu, Hongtao; Chen, Chunbo; Zheng, Hongwei; Luo, Geping; Yang, Liao; Wang, Weisheng; Wu, Shixin; Ding, Jianli
• Is Part Of: International journal of remote sensing. Volume 41:Issue 12(2020)
• Page Start: 4470
• Page End: 4495
• Publication Date: 2020-06-17
• Subjects: Remote sensing -- Periodicals; Télédétection -- Périodiques; 621.3678
• DOI: 10.1080/01431161.2020.1718239
• Languages: English
• ISSNs: 0143-1161

Abstract

ABSTRACT: Salinization is a major eco-environmental threat in arid and semi-arid regions. Machine learning combined with salinization-related factors extracted from satellite images and Digital Elevation Model (DEM) data to monitor salinization is popular in recent years. The hyper-parameters referring to the parameters prior to fitting the model to the data and the features referring to the factors used for model establishment are vital to modelling accuracy, while the optimization of the above two was not given enough attention. In this study, we proposed a novel approach for simultaneously identifying the input features and hyper-parameters of Support Vector Regression (SVR) based on Adaptive Genetic Algorithm (AGA) for quantitative assessment of salinization in Weigan-Kuqa river delta oasis (Wei-Ku oasis), Sangong River Basin, and Qitai oasis of Xinjiang. First, a total of 41 salinization-related factors of 7 categories were extracted from Landsat 5 TM and DEM data. In each sub-region, the Pearson's correlation analysis was developed between Soil Salt Content (SSC) and salinization-related factors, and the factors significantly correlated with SSC were arranged in descending order of absolute correlation coefficient to form the Candidate Feature Variables (CFVs). The ration of Coefficient of Determination ( R 2 ) and Root Mean Square Error (RMSE) multiplied by 1000 was considered as the fitness function of Genetic Algorithm (GA) and AGA. The CFVs and hyper-parameters were combined together and binary coded, then brought into the AGA and GA for simultaneous feature selection and hyper-parameters optimization of SVR, and established salinization monitoring models (AGA-SVR, GA-SVR). In order to highlight the importance of identifying the input features and hyper-parameter to modelling accuracy, the GS-SVR with all the CFVs as input was established using Grid Search (GS) algorithm to optimize the hyper-parameters. Finally, the salinization maps predicted by three models were compared. The results showed that the sensitivity of salinization-related factors to SSC varied with regions, and 25, 16, 24 salinization-related factors were selected as CFVs in Wei-Ku oasis, Sangong River Basin, and Qitai oasis, respectively. Compared with GS-SVR, the GA-SVR and AGA-SVR got more accurate salinization monitoring with fewer features as input, and fitness generated by AGA-SVR increased by 25.968% in Wei-Ku oasis, 25.159% in Sangong River Basin, 27.568% in Qitai oasis, respectively. Both nature and human factors lead to salinization. The difference of Land Surface Temperature (LST) was the main contributor of different salinization between Wei-Ku oasis of southern Xinjiang and the other two sub-regions of northern Xinjiang. The differences in soil texture, irrigation methods, and livestock carrying capacity were the main factors resulting in differences of salinization in Qitai oasis and Sangong River Basin. Our study shows that the proposed approach can provide technical support for accurate salinization monitoring.