Gholami Hamid, Mohammadifar Aliakbar, Song Yougui, Li Yue, Rahmani Paria, Kaskaoutis Dimitris G, Panagos Panos, Borrelli Pasquale
Department of Natural Resources Engineering, University of Hormozgan, Bandar-Abbas, Hormozgan, Iran.
State Key Laboratory of Loess and Quaternary Geology, Institute of Earth Environment, Chinese Academy of Sciences, Xi'an, 710061, China.
Sci Rep. 2024 Aug 15;14(1):18951. doi: 10.1038/s41598-024-70125-y.
Spatial accurate mapping of land susceptibility to wind erosion is necessary to mitigate its destructive consequences. In this research, for the first time, we developed a novel methodology based on deep learning (DL) and active learning (AL) models, their combination (e.g., recurrent neural network (RNN), RNN-AL, gated recurrent units (GRU), and GRU-AL) and three interpretation techniques (e.g., synergy matrix, SHapley Additive exPlanations (SHAP) decision plot, and accumulated local effects (ALE) plot) to map global land susceptibility to wind erosion. In this respect, 13 variables were explored as controlling factors to wind erosion, and eight of them (e.g., wind speed, topsoil carbon content, topsoil clay content, elevation, topsoil gravel fragment, precipitation, topsoil sand content and soil moisture) were selected as important factors via the Harris Hawk Optimization (HHO) feature selection algorithm. The four models were applied to map land susceptibility to wind erosion, and their performance was assessed by three measures consisting of area under of receiver operating characteristic (AUROC) curve, cumulative gain and Kolmogorov Smirnov (KS) statistic plots. The results revealed that GRU-AL model was considered as the most accurate, revealing that 38.5%, 12.6%, 10.3%, 12.5% and 26.1% of the global lands are grouped at very low, low, moderate, high and very high susceptibility classes to wind erosion hazard, respectively. Interpretation techniques were applied to interpret the contribution and impact of the eight input variables on the model's output. Synergy plot revealed that the soil carbon content exhibited high synergy with DEM and soil moisture on the model's predictions. ALE plot showed that soil carbon content and precipitation had negative feedback on the prediction of land susceptibility to wind erosion. Based on SHAP decision plot, soil moisture and DEM presented the highest contribution on the model's output. Results highlighted new regions at high latitudes (southern Greenland coast, hotspots in Alaska and Siberia), which exhibited high and very high land susceptibility to wind erosion.
对土地风蚀敏感性进行空间精确制图对于减轻其破坏性后果至关重要。在本研究中，我们首次开发了一种基于深度学习（DL）和主动学习（AL）模型、它们的组合（例如循环神经网络（RNN）、RNN-AL、门控循环单元（GRU）和GRU-AL）以及三种解释技术（例如协同矩阵、SHapley值相加解释（SHAP）决策图和累积局部效应（ALE）图）的新方法，以绘制全球土地风蚀敏感性图。在这方面，研究了13个变量作为风蚀的控制因素，并通过哈里斯鹰优化（HHO）特征选择算法选择了其中8个（例如风速、表土碳含量、表土粘土含量、海拔、表土砾石碎片、降水量、表土砂含量和土壤湿度）作为重要因素。应用这四种模型绘制土地风蚀敏感性图，并通过由接收器操作特征（AUROC）曲线下面积、累积增益和柯尔莫哥洛夫-斯米尔诺夫（KS）统计图组成的三种度量来评估它们的性能。结果表明，GRU-AL模型被认为是最准确的，显示全球土地分别有38.5%、12.6%、10.3%、12.5%和26.1%被归类为对风蚀危害的极低、低、中、高和极高敏感性类别。应用解释技术来解释八个输入变量对模型输出的贡献和影响。协同图显示，土壤碳含量在模型预测方面与数字高程模型（DEM）和土壤湿度表现出高度协同。ALE图表明，土壤碳含量和降水量对土地风蚀敏感性预测有负反馈。基于SHAP决策图，土壤湿度和DEM对模型输出的贡献最大。结果突出了高纬度地区的新区域（格陵兰岛南部海岸、阿拉斯加和西伯利亚的热点地区），这些地区对风蚀表现出高和极高的土地敏感性。
