Classification of landslides on the southeastern Tibet Plateau based on transfer learning and limited labelled datasets

Abstract

ABSTRACT Rainfall and freeze-thaw landslides are common occurrences on southeastern Tibet Plateau; however, it can be difficult to categorize them via field surveys. Landslide classification can be determined using image processing methods, such as deep learning. However, the performance of classification is restricted by limited datasets resulting from small numbers of landslides. A new deep learning method is proposed herein for classifying landslides by remote sensing images using the VGG-19 and transfer learning to compensate for an insufficient number of labelled samples. Transfer learning was used to fine tune the classification model. The results of ablation experiments show that the classification model combined with transfer learning can obtain a near zero total loss value in the training process. The proposed method can achieve a lower loss value than direct learning and has a superior generalization ability. The prediction accuracies, Kappa index and F1-score of VGG-19 when using transfer learning and 30 training samples reached 98%, 0.979 and 0.982 respectively. The experiments indicate that the proposed method can be applied to landslide classification on plateaus.