Identification of geo-environmental factors on Benggang susceptibility and its spatial modelling using comparative data-driven methods

• Published in: Soil & tillage research Vol. 208; p. 104857
• Main Authors: Wei, Yujie, Wu, Xinliang, Wang, Junguang, Yu, Hongliang, Xia, Jinwen, Deng, Yusong, Zhang, Yong, Xiang, Yu, Cai, Chongfa, Guo, Zhonglu
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.04.2021
• Subjects: Benggang; Multilayer perceptron; Multinomial logistic regression; Random forest; Susceptibility modelling; Multinomial logistic regression; Multilayer perceptron; Random forest; Benggang; Susceptibility modelling
• ISSN: 0167-1987; 1879-3444
• DOI: 10.1016/j.still.2020.104857

[Display omitted] •Meteorology was the most important driving factor, followed by topography.•Rainfall erosivity had the highest importance to Benggang erosion in Fujian province.•The critical range of rainfall erosivity to Benggang is 7475 ∼ 8349 MJ mm(ha ha)−1.•Random forest was suitable for assessment of the regional Benggang susceptibility.•This study provides a methodological framework to Benggang susceptibility prediction. Benggang, a fragmented erosion landform, is the most serious threat to the sustainable development of ecosystem and economy in southern China. However, few studies have been conducted on variations and conditions of the geo-environmental factors that influence the susceptibility of Benggang. To address this challenge, the spatial susceptibility of Benggang was investigated using three data-driven methods (multinomial logistic regression, MLR; random forest, RF; multilayer perceptron, MLP) in Fujian province, and the dominant driving forces to Benggang susceptibility were identified by Boruta algorithm, while the contribution of each factor’s class to Benggang occurrence was determined by frequency ratio (FR). Herein, thirteen conditioning factors with relate to geomorphology, climate, vegetation, and land use was employed. Among these factors, the most noticeable contributions to Benggang occurrence at a provincial scale were from rainfall erosivity, soil moisture, and NDVI with their corresponding critical ranges of 7475 ∼ 8349 MJ mm(ha ha)−1, 0.268 ∼ 0.297 m3 m-3, and 0.58 ∼ 0.68. According to the area under the success rate and prediction curves (AUC) for the evaluation of susceptibility modelling, the averaged training accuracies for susceptibility levels were 85.70 %, 99.99 % and 99.23 %, while the prediction accuracies were 85.79 %, 99.17 % and 97.07 % for MLR, RF and MLP, respectively. In general, RF with its AUC value, exceeding 98% at each susceptibility level for both success and prediction curve, was especially suitable for the assessment of the regional Benggang susceptibility. Furthermore, the methodological framework of this study could be implanted in Benggang erosion regions with similar conditions and data availabilities, whereas different data mining techniques and condition factors should be considered at different scales in our future work.