Full-Scale Feature Pyramid Information Propagation for Postearthquake Landslide Detection

Accurate detection of post-earthquake landslides from high-resolution Earth observation imagery is critical for emergency response and hazard assessment, yet remains challenging due to extreme scale variability and complex terrain backgrounds. Landslides in post-seismic scenes range from small, fragmented failures to large, continuous slope collapses, posing significant difficulties for reliable multi-scale representation. To address these challenges, this study proposes a task-driven deep learning framework that emphasizes full-scale information propagation across feature pyramid levels (P2–P5) together with efficient global contextual modeling for post-earthquake landslide detection. The proposed framework enables persistent and bidirectional feature interaction within the feature pyramid, allowing fine-grained spatial details from shallow layers and high-level semantic context from deep layers to be jointly propagated and refined. In addition, a magnitude-aware linear attention mechanism is incorporated to enhance long-range contextual dependencies with improved computational efficiency, making the framework suitable for high-resolution, large-area remote sensing imagery. The proposed method is evaluated on two post-earthquake landslide datasets with different scene complexities and scale distributions. Experimental results demonstrate that the proposed framework consistently outperforms representative convolutional- and transformer-based detectors, achieving mean average precision (mAP) values of 0.937 and 0.882 on the two datasets, respectively. Qualitative analyses further indicate improved localization accuracy and reduced false alarms, particularly for small-scale and elongated landslides under complex background conditions. These results highlight the importance of explicitly modeling full-scale feature pyramid information propagation for robust post-earthquake landslide detection and demonstrate the practical potential of the proposed framework for large-scale post-disaster landslide mapping.