Sensitivity analysis of regional rainfall-induced landslide based on UAV photogrammetry and LSTM neural network

• Published in: Journal of mountain science Vol. 20; no. 11; pp. 3312 - 3326
• Main Authors: Zhao, Lian-heng, Xu, Xin, Lyu, Guo-shun, Huang, Dong-liang, Liu, Min, Chen, Qi-min
• Format: Journal Article
• Language: English
• Published: Heidelberg Science Press 01.11.2023; Springer Nature B.V
• Subjects: Aerial photography; Analysis; Boreholes; Earth and Environmental Science; Earth Sciences; Ecology; Empirical analysis; Environment; Geography; Inventories; Landslides; Landslides & mudslides; Long short-term memory; Neural networks; Original Article; Photogrammetry; Plateaus; Pore water; Precipitation; Rainfall; Rainfall infiltration; Rainfall intensity; Records; Sensitivity analysis; Slope stability; Soil properties; Unmanned aerial vehicles; UAV photography; TRIGRS; Rainfall landslide; Regional landslide; LSTM neural network
• ISSN: 1672-6316; 1993-0321; 1008-2786
• DOI: 10.1007/s11629-023-7991-z

Rainfall stands out as a critical trigger for landslides, particularly given the intense summer rainfall experienced in Zheduotang, a transitional zone from the southwest edge of Sichuan Basin to Qinghai Tibet Plateau. This area is characterized by adverse geological conditions such as rock piles, debris slopes and unstable slopes. Furthermore, due to the absence of historical rainfall records and landslide inventories, empirical methods are not applicable for the analysis of rainfall-induced landslides. Thus we employ a physically based landslide susceptibility analysis model by using high-precision unmanned aerial vehicle (UAV) photogrammetry, field boreholes and long short term memory (LSTM) neural network to obtain regional topography, soil properties, and rainfall parameters. We applied the Transient Rainfall Infiltration and Grid-Based Regional Slope-Stability (TRIGRS) model to simulate the distribution of shallow landslides and variations in porewater pressure across the region under different rainfall intensities and three rainfall patterns (advanced, uniform, and delayed). The landslides caused by advanced rainfall pattern mostly occurred in the first 12 hours, but the landslides caused by delayed rainfall pattern mostly occurred in the last 12 hours. However, all the three rainfall patterns yielded landslide susceptibility zones categorized as high (1.16%), medium (8.06%), and low (90.78%). Furthermore, total precipitation with a rainfall intensity of 35 mm/h for 1 hour was less than that with a rainfall intensity of 1.775 mm/h for 24 hours, but the areas with high and medium susceptibility increased by 3.1%. This study combines UAV photogrammetry and LSTM neural networks to obtain more accurate input data for the TRIGRS model, offering an effective approach for predicting rainfall-induced shallow landslides in regions lacking historical rainfall records and landslide inventories.