江西省龙南县滑坡易发性评价

区域滑坡易发性评价是国内外地质灾害研究的重点和热点。目前,国内外学者已提出了支持向量机(support vector machine,SVM)、BP神经网络和随机森林等多种模型并成功用于滑坡易发性评价。但在利用这些机器学习模型评价滑坡易发性时,存在着参数选取困难、建模效率低、模型训练时间长和对评价指标解释能力弱等问题。为简化建模过程、提高预测精度及增强模型的可解释性,提出了基于频率比分析和偏最小二乘回归法(partial least squares regression,PLSR)的滑坡易发性评价模型。PLSR模型很好地发挥了主成分分析和回归分析的优势,考虑了评价指标间的内在联系,具有建模过程简洁、可解释性强的优点。将结合频率比法的PLSR模型应用于江西省龙南县滑坡易发性评价,并与BP神经网络、SVM模型的易发性评价结果进行对比。研究表明:PLSR模型的预测精度优于BP神经网络,且与SVM模型预测精度接近;另外,在综合考虑建模效率、预测精度和模型解释能力的情况下,PLSR模型具有更高的实用性。

Regional Landslide Susceptibility Assessment based on Frequency Ratio analysis and Partial Least Squares Regression model

Landslide susceptibility assessment (LSA) is one of the research hotspots of geological disasters around the world, many scholars have put forward machine learning models such as Support Vector Machine (SVM), BP Neural Network (BPNN) and Random Forest to do LSA and these models have been successfully applied for LSA. However, there are some problems, such as difficult selection of parameters, low efficiency of modeling, long model training time and low explanatory power, when these machine learning models are used to deal with LSA. In order to simplify the modeling process, to improve the prediction accuracy and to enhance the model interpretability, a LSA model combining frequency ratio analysis and Partial Least Squares Regression (PLSR) method is proposed in this study. The PLSR model takes advantages of principal component analysis and regression analysis, and it can effectively reveal the inherent relationships between landslides and landslide-related environmental. Furthermore, PLSR model has the advantages of concise modeling processes and strong model interpretation capability. In this study, the frequency ratio based PLSR model is applied to do LSA of Longnan County, Jiangxi Province. Meanwhile, LSA results of PLSR model are compared with those of BP neural network and SVM model. Results show that, the prediction accuracy of PLSR model is higher than that of BP neural network, and is close to that of SVM model. Results also show that the PLSR model is efficient for LSA with high modeling efficiency, high prediction accuracy and good model interpretation ability.