Landslide Susceptibility Assessment of Lixian County, Gansu Province, China Using a Physics‐Integrated Multi‐Task Neural Network

ABSTRACT

Landslide susceptibility assessment (LSA) is a critical strategy for disaster prevention and mitigation. However, current machine learning‐based LSA algorithms often exhibit instability under data‐sparse conditions and suffer from a lack of physical interpretability. To address these challenges, this study proposes a physics‐integrated multi‐task neural network (PIMTN). The proposed framework incorporates slope‐stability information derived from the infinite slope model into a multi‐task neural network through an auxiliary learning task. The model simultaneously performs landslide susceptibility classification and learns a physics‐derived diluted probability of failure (DPOF). Through joint optimization, the auxiliary DPOF task influences the shared representation and guides susceptibility prediction towards spatial patterns consistent with physics‐derived slope‐stability information. Comparative experiments demonstrate that the PIMTN outperforms benchmark models (Precision: 0.737, F 1‐Score: 0.754, AUC: 0.817). Furthermore, quantitative analysis confirms the superiority of the proposed model and the effectiveness of knowledge transfer from the physical domain. This framework presents a novel solution for enhancing model interpretability and improving predictive robustness under limited‐inventory conditions.