基于耦合信息量法选择负样本的区域滑坡易发性预测

机器学习(Machine Learning, ML)模型预测滑坡易发性时选择合理的负样本对预测结果具有重要影响，现有研究大多从整个研究区或从低坡度等特定属性区内随机选择负样本，这些负样本往往不够准确或以偏概全，降低了易发性制图的可靠性。为解决这一问题，拟提出耦合信息量法(Information Value, IV)的ML模型开展易发性建模。以江西省瑞金市为例，采用IV法将环境因子的属性值转化为对滑坡贡献的信息量值，划定极低和低易发区并从中随机选择出ML模型训练验证用的负样本数据，构建全新的信息量-支持向量机(IV-SVM)、信息量-随机森林(IV-RF)耦合模型并预测瑞金滑坡易发性；进一步与从全区随机选择负样本的单独SVM和RF模型，以及从坡度小于2°的特定属性区内随机选负样本的低坡度SVM和RF模型做对比研究；最后采用Kappa系数和ROC曲线等指标验证和比较建模结果。IV-SVM和IV-RF模型的Kappa系数为0.828和0.9146且对应ROC曲线的AUC值为0.876和0.939，分别高于单独SVM、RF和低坡度SVM、RF模型；同时，IV-SVM和IV-RF模型的易发性概率分布的平均值较小而标准差较大。结果表明：1) IV-SVM和IV-RF模型具有比单独SVM和RF模型，以及低坡度SVM和RF模型更高的滑坡易发性预测精度且更有效的反映了瑞金滑坡易发性分布规律；2) RF模型相较于SVM模型具有更高的预测精度；3) IV-RF等耦合模型能够弥补单独模型存在的负样本采样不准确和低坡度模型对坡度因子区间选择的缺点而提高预测精度更加合适机器学习的滑坡易发性预测建模。总之，本文研究为机器学习预测滑坡易发性的负样本采样方法提供新思路。

Regional Landslide Susceptibility Prediction Based on Negative Sample Selected by Coupling Information Value Method

For the landslide susceptibility prediction (LSP) based on machine learning (ML) models, the reasonable selection of negative samples has an important influence on the LSP performance. Generally, the main selection methods include randomly selecting from the whole study area or from specific attribute areas such as low slope. The negative samples selected by the above methods are often inaccurate or biased, resulting in low accuracy and low reliability of LSP. To solve this problem, the coupling model of ML and Information Value (IV) method is proposed for LSP. Taking Ruijin City as the study area, the attribute values of the environmental factors are transformed into the IV values of the contribution to the landslide to obtain the very low and low susceptibility areas. The negative samples are randomly selected in the above areas for training and validation of machine learning models. The new coupling models of IV-SVM and IV-RF were constructed for LSP of Ruijin. Further, IV-SVM and IV-RF models are compared with the single SVM and RF model with negative samples randomly selected from the whole study area, as well as the low-slope SVM and RF model with negative samples randomly selected from specific attribute areas with a slope less than 2°. Finally, Kappa Coefficient (KC) and ROC curve are used to verify and compare the modeling results. The AUC values of ROC curve and KC of IV-SVM and IV-RF models were 0.828, 0.920 and 0.876, 0.988, which were higher than those of single SVM, RF model and low-slope SVM, RF model, respectively. Meanwhile, IV-SVM and IV-RF models have smaller mean value and larger standard deviation of susceptibility probability distribution. Results show that: 1) IV-SVM and IV-RF models have the higher LSP accuracies than those of the single SVM, RF model and low-slope SVM, RF model, respectively; 2) RF model has higher LSP accuracy comparing to the SVM model; 3) The coupling model such as IV-RF can address the inaccuracy of negative sample sampling existing in the single model and the shortcomings of the low slope model in the selection of slope interval, thus improving the LSP accuracy. In conclusion, this study provides a new idea for the negative sample sampling method for LSP using ML models.