Landslide detection based on pixel-level contrastive learning for semi-supervised semantic segmentation in wide areas

In response to the challenge of insufficiently labeled landslide samples, which hinders the development of supervised deep learning for landslide detection, this study presents a semi-supervised semantic segmentation model based on pixel-level contrastive learning. This method effectively uses the rich semantic and feature information in a vast pool of unlabeled samples, even when the number of labeled samples is limited, thereby achieving accurate landslide detection. In this study, we employed virtual adversarial training (VAT) to augment the model’s robustness against the most sensitive directions in the data space, effectively mitigating the insufficient model generalizability caused by the diversity in landslide morphology and surface cover. In addition, this study innovatively adopted a pixel-level contrastive learning mechanism that is constrained by probability, thereby establishing a highly structured pixel feature space. This approach facilitates the aggregation of intraclass pixels and the dispersion of interclass pixels, thus enhancing the precision and reliability of the model for detecting various types of landslides. To assess the accuracy and reliability of the model, we constructed a dataset comprising 1022 labeled samples and 2776 unlabeled samples of landslides from Planet optical imagery and digital elevation model (DEM) data within a geographical area spanning 1076.2 km in length and 254.4 km in width and ranging from western Sichuan to eastern Tibet, China. Comparative validation was conducted against existing semi-supervised methods, namely the error localization network and SemiSeg-Contrastive, as well as fully supervised approaches including SegFormer, the Swin-transformer, and Fast SCNN. The experimental results indicated that our model proficiently detected landslides, detecting 432 landslides in the study area with a detection accuracy of 0.913, an accuracy of 0.928 on the validation set, an F1 score of 0.91, and a mean intersection over union (mIoU) of 84.2%. These results surpassed those of the compared models, confirming the reliability and effectiveness of our method. The achievements of this research can serve as a reference and guide for deep-learning-based intelligent landslide detection.