Advanced deep learning–based approaches for semantic segmentation in precise landslide detection and susceptibility assessment

Landslides are a formidable natural geological hazard that impose significant disruptions on societal and economic functions. The magnitude of impact of landslides on human life makes early detection imperative in landslide-prone areas. The aims of this study were to detect landslide locations triggered by the Taiwan Morakot typhoon in 2009 using Sentinel-2 imagery. The fully convolutional network; image cascade network; red, green, blue plus depth (RGB-D) fusion network (RDFNet); and segmented neural network algorithms were developed to more accurately detected historical landslide locations and afterward generate a landslide susceptibility map using standard and optimized deep learning models (convolutional neural network [CNN], CNN-whale optimization algorithm, and CNN-Harris hawk optimization [HHO]). Analysis using frequency ratio to evaluate the relationship between classes of different causative factors and landslide occurrence revealed that slope length and land cover were most strongly correlated with landslide occurrence. Among the algorithms tested, RDFNet demonstrated superior accuracy in detecting historical landslide locations, achieving the highest F1-score (0.67), precision (0.55), recall (0.84), and accuracy (0.92). The landslide inventory dataset derived from RDFNet was used in landslide susceptibility modeling, and was divided into 70 % and 30 %, respectively, for training and testing the CNN-based models. Mean square error ( MSE ), root mean square error ( RMSE ), standard deviation ( StD ), and area under the receiver operating characteristic curve ( AUROC ) were used to evaluate goodness-of-fit and predictive ability of the CNN-based models. The most reliable outcome was obtained using CNN-HHO, with AUROC = 0.85, MSE = 0.017, RMSE = 0.132, and StD = 0.129 during the testing step. By leveraging insights provided by modeling into areas with high landslide susceptibility, authorities can proactively implement measures to mitigate the potential consequences of landslides and promote sustainable stewardship of natural resources. • Strong correlation found between landslide occurrence and changes in land cover. • RDFNet algorithm showed superior accuracy in detecting landslide locations. • Optimized CNN-based models successfully generated landslide susceptibility maps. • CNN-HHO outperformed the other models in detecting landslide-prone areas.