NL-YOLOv5: a model with a larger receptive field and the ability to globally acquire features

Landslide disasters cause severe casualties and economic losses, demanding rapid and accurate detection from high-resolution remote sensing imagery. Traditional methods struggle with insufficient landslide samples and low detection accuracy.

To address this, we propose a dual solution: (1) DCGAN-based data augmentation generating 3,429 synthetic landslide samples from Google Earth imagery, significantly improving model generalization; (2) an NL-YOLOv5 model integrating non-local attention (NLA) and an improved LK-SPP module (based on large-kernel convolution concepts) to enhance global information capture.

The NL-YOLOv5 model achieves 80% detection accuracy (a 5% improvement over baseline), with 7% higher precision, 2% higher recall, and 5% higher F1-score, while maintaining real-time speed at 69 f/s.

This work delivers a practical solution for high-precision landslide detection in remote sensing applications.