An adaptive feature extraction lightweight network for enhanced landslide detection

Rapid detection of landslides is essential for timely disaster warning and minimizing potential losses. However, identifying landslides in natural environments remains challenging due to their irregular morphologies, scale variability, and the high computational demands of deep neural networks. To address these challenges, this study proposes the Adaptive Feature Extraction Lightweight Network for Landslide Detection (AFL-NET). The backbone network incorporates the Ghostv2C2f module, which integrates ghost feature generation and dynamic spatial attention to reduce redundant computation and parameter count while preserving spatial detail. For feature fusion, the SimAM-Augmented Bi-directional Feature Pyramid Network (SBIFPN) is introduced to enhance multi-scale representation via adaptive weight normalization and energy-guided attention, improving the model's response to landslides of varying shapes and scales. At the detection stage, the Spatial Modeling and Context-aware C2f (SMC2f) module combines lightweight contextual encoding with multi-head self-attention to suppress background interference and strengthen structural reasoning in complex terrain. Experimental results show that AFL-NET achieves a 3.9% improvement in mAP over the baseline YOLOv11n, reaching 90.1%, while reducing parameter by 15.3%. These improvements make AFL-NET highly suitable for real-time, on-device deployment in disaster monitoring applications, offering reliable technical support for early warning and risk mitigation.