Domain-Adaptive and Efficient Neural Network for Accelerated Post-Earthquake Landslide Recognition

Abstract

Co-seismic landslides, triggered by major earthquakes, are sudden, widespread events that necessitate immediate and accurate information for disaster response. Conventional deep learning (DL) methods used for remote sensing image interpretation often suffer from complex model architecture, slow inference speeds, and poor generalization across diverse geographical scenes, rendering them inadequate for rapid post-disaster assessment. This paper introduces a novel Lightweight Deep Transfer Learning (LDTL) framework specifically designed to accelerate cross-scene landslide identification. The core of this framework is an efficient Multi-scale Feature Fusion Lightweight Network (MSF-LiteNet) optimized for computational efficiency and parameter reduction. By integrating a progressive knowledge transfer strategy, the MSF-LiteNet can rapidly leverage prior landslide feature knowledge from multiple source domains and perform robust, precise adaptation to new disaster zones. Empirical validation across multiple typical seismic disaster scenes (including events in Japan and China) demonstrates that the proposed LDTL framework significantly boosts deployment speed and cross-domain detection capability while maintaining high recognition accuracy, offering a viable technical route for real-time global geohazard assessment.