Early identification of potential loess landslide using convolutional neural networks with skip connection: a case study in northwest Lvliang City, Shanxi Province, China

Loess landslide is one of the most harmful and serious geological hazards in the Loess Plateau of China. Early identification of potential loess landslide is an urgent need for its prevention. Traditional methods, e.g. support vector machines and decision trees, often suffer complicated data pre-processing, multitudinous causative factors, or low accuracy. This study aims to develop a high-performance loess landslide early identification model based on convolutional neural networks (CNNs). A case study was carried out in northwest Lvliang, China, where loess landslide is a major concern. Two hundred and six loess landslide cases were interpreted by comparing remote sensing images of two time phases, and were randomly divided into a training set (80%; 165) and a validation set (20%; 41). Four algorithms were developed, including a CNN structure with skip connection using data with (S-C) or without (S-N) slope crest and plain CNN structure using data with (P-C) or without (P-N) slope crest. The results show that the S-C structure is the most suitable for early identification of potential loess landslides because it achieved the highest overall accuracy (OA = 0.902) and largest area under the receiver operating characteristic curve (AUC = 0.932) on the validation set.

Automated summary

This study aims to develop a high-performance early identification model for loess landslides based on convolutional neural networks (CNNs). By comparing remote sensing images and analyzing data, it is found that the CNN structure with slope crest data is the most suitable for early identification of potential loess landslides.