Landsat data-driven identification of surface soil freeze and thaw states: a comparison of various target detection and semantic segmentation models

Frozen soil is a vital component of the cryosphere and is sensitive to climate changes, accelerating thaw when the anthropogenic climate warms. Identifying the state of soil freeze/thaw is crucial for effective water resource management, cold region engineering construction, agriculture development, etc. In this study, we used Landsat remote sensing images data to identify the freeze/thaw state of frozen soil, focusing on the eastern Tibetan Plateau (ETP) as a case study. We compared the target detection models of two architectures and the semantic segmentation model framework of four models with different encoders. We used Mean Pixel Accuracy (MAP), Mean Intersection over Union (MIoU), Taylor diagrams, and surface borehole data as metrics to analyze the model’s accuracy. Our results showed that the YOLOv8-Large target detection model performed better than other detection models, with the MAP50 of 0.53. The ResNet34 encoder in the semantic segmentation model demonstrated the highest accuracy of 86.92%, effectively resisting interference, and could identify freeze–thaw with higher accuracy in noisy images. Our study provides a high-resolution pathway to identify freeze/thaw state of frozen soils, which can benefit studies in cryosphere science, hydrology, ecology and climate change.