Machine learning analysis demonstrates improved landslide detection speed and accuracy using augmented remote sensing imagery, indicating high utility for lightweight disaster monitoring.
Key Points
Develop a lightweight, high-accuracy deep learning framework to improve landslide detection in remote sensing imagery with limited training data.
Collected high-quality remote sensing imagery from southeastern Xizang to establish a landslide dataset.
Designed Landslide-DCGAN (L-DCGAN) with efficient channel attention (ECA) in the generator and squeeze-and-excitation (SE) modules in the discriminator for synthetic data augmentation.
Developed Landslide-YOLOv8 (L-YOLOv8) by integrating cross-stage partial pyramid convolution (CSPPC), grouped and shuffled convolution (GSConv), and spatial and channel synergistic attention (SCSA).
Reduced total model parameters by approximately 41.78% compared to the baseline YOLOv8 architecture.
Increased average detection accuracy on the landslide sample dataset by 0.57%.
Improved inference speed with a 33.87% increase in frames per second (FPS).