Improved Landslide Detection in Southeastern Xizang Based on L-DCGAN-Augmented Data and the L-YOLOv8 Model

Abstract

Landslides are common geological disasters that cause extensive damage and significant harm. Efficient recognition of landslide remote sensing images holds both theoretical and practical importance. However, practical landslide detection faces challenges including limited data, high computation, large memory use, and low accuracy. In this study, high-quality landslide remote sensing images were collected from southeastern Xizang to construct a landslide remote sensing data set. To address the issue of insufficient samples, an improved model, landslide-deep convolutional generative adversarial network (L-DCGAN), based on DCGAN, was proposed to enhance landslide remote sensing data. The enhanced data set was constructed by incorporating efficient channel attention (ECA) and squeeze-and-excitation (SE) modules into the generator and discriminator, respectively. Additionally, an improved YOLOv8 model, landslide-you only look once version 8(L-YOLOv8), was proposed. First, the cross-stage partial pyramid convolution (CSPPC) and grouped and shuffled convolution (GSConv) modules replaced the cross stage partial layer with 2 branches (C2f) module and ordinary convolution in the network to reduce the number of model parameters and computational complexity. Then, the spatial and channel synergistic attention (SCSA) mechanism was introduced to improve detection accuracy. Experimental results showed that compared with the original YOLOv8 algorithm, the number of model parameters in L-YOLOv8 was reduced by approximately 41.78%, and the average accuracy on the landslide sample data set was increased by 0.57%, and frames per second (FPS) was improved by 33.87%. The model achieved lightweight design while improving detection speed and accuracy, and demonstrated superior performance on the landslide remote sensing data set.