基于深度学习的滑坡灾害易发性分析

滑坡灾害成因机理复杂、影响因素众多,深度学习作为当前人工智能领域的热点,能够更好地模拟滑坡灾害的形成并准确预测潜在的斜坡。为了挖掘深度学习在滑坡易发性的应用潜能,本文构建了一维、二维和三维的滑坡数据表达形式,并提出3种基于卷积神经网络模型（Convolutional Neural Networks, CNN）的滑坡易发性分析处理框架：基于CNN分类器、基于CNN与逻辑回归的融合和基于CNN集成,最后以江西省铅山县为研究对象进行验证,结果表明：所有基于CNN的易发性模型都能够获得准确且可靠的滑坡易发性分析结果。其中,基于二维数据的CNN模型在所有单分类器中预测精度最高,为78.95%。此外,二维CNN特征提取能够显著提升逻辑回归的预测精度,其准确率提升7.9%。最后,异质集成策略能够大幅度提升基于CNN分类器的滑坡预测精度,其准确率提升4.35%~8.78%。

Landslide Susceptibility Analysis based on Deep Learning

The formation mechanism of landslide disasters is complicated and there are many influencing factors. It is imperative to explore a low-cost and highly applicable method to manage and prevent landslide disasters. As a hot spot in the current artificial intelligence field, deep learning can better simulate the formation of landslide disasters and accurately predict potential slopes. Thus, to explore the application potential of deep learning, this paper constructs one-dimensional, two-dimensional, and three-dimensional forms of landslide data, and then introduces three Convolutional Neural Networks (CNN)-based landslide susceptibility analysis frameworks, including CNN-based classifiers, integrated models, and ensemble models. The proposed deep learning methods were applied to Yanshan County, Jiangxi Province for experiments. 16 landslide influencing factors were first selected for modelling based on the geomorphological, hydrological, and geological environment conditions of the study area. These factors include altitude, aspect, distance to faults, land use, lithology, normalized difference vegetation index, plan curvature, profile curvature, rainfall, distance to rivers, distance to roads, slope, soil, stream power index, sediment transport index, and topographic wetness index. Then, the multi-collinearity analysis and relief-F algorithm were used to analyze and screen the influencing factors. All CNN-based methods were constructed and validated based on several statistical measures of accuracy, root mean square error, mean absolute error, sensitivity, specificity, and the receiver operation characteristic curve. Finally, the susceptibility value of each pixel in the study area was predicted based on the CNN-based methods, and the entire study areas were reclassified into five susceptibility categories: very low, low, moderate, high, and very high. The factor analysis results show that the plan curvature, profile curvature, stream power index, and sediment transport index are redundant factors and should be removed from further modelling process. The model evaluation results demonstrate that all CNN-based models can obtain accurate and reliable landslide susceptibility mapping results. The two-dimensional CNN model achieved the highest prediction accuracy of 78.95% among single CNN models. Moreover, the performance of logistic regression was effectively improved by combining the two-dimensional CNN for feature extraction, with an accuracy improvement of 7.9%. Besides, the heterogeneous ensemble strategy can greatly improve landslide prediction accuracy when using CNN classifiers, with an accuracy improvement between 4.35% and 8.78%. Generally, the CNN has been proven to have huge application potential in landslide susceptibility analysis and can be implemented in other landslide-prone areas with similar geo-environmental conditions.