Department of Geography, University of Gour Banga, Malda, West Bengal, India.
College of Marine Sciences and Engineering, Nanjing Normal University, Nanjing 210023, China; Department of Natural Resources and Environmental Engineering, College of Agriculture, Shiraz University, Shiraz, Iran.
Sci Total Environ. 2019 Jun 10;668:124-138. doi: 10.1016/j.scitotenv.2019.02.436. Epub 2019 Mar 1.
Gully erosion is one of the most effective drivers of sediment removal and runoff from highland areas to valley floors and stable channels, where continued off-site effects of water erosion occur. Gully initiation and development is a natural process that greatly impacts natural resources, agricultural activities, and environmental quality as it promotes land and water degradation, ecosystem disruption, and intensification of hazards. In this research, an attempt is made to produce gully erosion susceptibility maps for the management of hazard-prone areas in the Pathro River Basin of India using four well-known machine learning models, namely, multivariate additive regression splines (MARS), flexible discriminant analysis (FDA), random forest (RF), and support vector machine (SVM). To support this effort, observations from 174 gully erosion sites were made using field surveys. Of the 174 observations, 70% were randomly split into a training data set to build susceptibility models and the remaining 30% were used to validate the newly built models. Twelve gully erosion conditioning factors were assessed to find the areas most susceptible to gully erosion. The predisposing factors were slope gradient, altitude, plan curvature, slope aspect, land use, slope length (LS), topographical wetness index (TWI), drainage density, soil type, distance from the river, distance from the lineament, and distance from the road. Finally, the results from the four applied models were validated with the help of ROC (Receiver Operating Characteristics) curves. The AUC value for the RF model was calculated to be 96.2%, whereas for those for the FDA, MARS, and SVM models were 84.2%, 91.4%, and 88.3%, respectively. The AUC results indicated that the random forest model had the highest prediction accuracy, followed by the MARS, SVM, and FDA models. However, it could be concluded that all the machine learning models performed well according to their prediction accuracy. The produced GESMs can be very useful for land managers and policy makers as they can be used to initiate remedial measures and erosion hazard mitigation in prioritized areas.
沟蚀是将高地的泥沙和径流转运到河谷和稳定河道的最有效驱动力之一，在这些地方，水蚀的持续场外影响仍在继续。沟蚀的发生和发展是一个自然过程，它极大地影响了自然资源、农业活动和环境质量，因为它促进了土地和水的退化、生态系统的破坏以及灾害的加剧。在这项研究中，尝试使用四种著名的机器学习模型（多元加性回归样条（MARS）、灵活判别分析（FDA）、随机森林（RF）和支持向量机（SVM））为印度帕特罗河流域的易受灾地区制作沟蚀易感性图，以进行管理。为了支持这一努力，使用实地调查对 174 个沟蚀点进行了观测。在 174 个观测中，随机抽取 70%作为训练数据集来建立易感性模型，其余 30%用于验证新建立的模型。评估了 12 个沟蚀条件因素，以找到最容易发生沟蚀的区域。诱发因素包括坡度梯度、海拔、平面曲率、坡度方向、土地利用、坡度长度（LS）、地形湿润指数（TWI）、排水密度、土壤类型、距河流的距离、距线性构造的距离和距道路的距离。最后，借助 ROC（接收者操作特征）曲线对四种应用模型的结果进行了验证。随机森林模型的 AUC 值计算为 96.2%，而 FDA、MARS 和 SVM 模型的 AUC 值分别为 84.2%、91.4%和 88.3%。AUC 结果表明，随机森林模型具有最高的预测精度，其次是 MARS、SVM 和 FDA 模型。然而，可以得出结论，根据预测精度，所有机器学习模型都表现良好。生成的 GESMs 对土地管理者和决策者非常有用，因为它们可用于启动优先地区的补救措施和侵蚀危害缓解。
