作者: Li Zhu , Lianghao Huang , Linyu Fan , Jinsong Huang , Faming Huang
DOI: 10.3390/S20061576
关键词:
摘要: Landslide susceptibility prediction (LSP) modeling is an important and challenging problem. features are generally uncorrelated or nonlinearly correlated, resulting in limited LSP performance when leveraging conventional machine learning models. In this study, a deep-learning-based model using the long short-term memory (LSTM) recurrent neural network conditional random field (CRF) cascade-parallel form was proposed for making LSPs based on remote sensing (RS) images geographic information system (GIS). The RS main data sources of landslide-related environmental factors, GIS used to analyze, store, display spatial big data. LSTM-CRF consists frequency ratio values factors input layers, LSTM feature extraction hidden full connection classification CRF landslide/non-landslide state output layers. can extract from different layers merge them into concrete features. calculate energy relationship between two grid points, extracted further smoothed optimized. As case applied Shicheng County Jiangxi Province China. A total 2709 landslide cells were recorded non-landslide randomly selected study area. results show that, compared with existing traditional algorithms, such as multilayer perception, logistic regression, decision tree, had higher rate (positive predictive rate: 72.44%, negative 80%, 75.67%). conclusion, novel data-driven deep that overcomes limitations algorithms achieves promising LSPs.