Identification of bald patches in degraded alpine meadows by UAV-based remote sensing and deep learning

Abstract

Alpine meadow patchiness is the starting point and an important feature of the formation of ‘Black Beach’ on the Tibetan Plateau, which directly threatens regional ecological security and economic development. Therefore, exploring an efficient, rapid and accurate method for identifying bald patches in degraded alpine meadows is of great significance for the dynamic monitoring and rational utilization in the grassland region. In this study, high-resolution remote sensing image data of degraded alpine meadows was obtained using a low-altitude Unmanned Aerial Vehicle (UAV)-mounted imager. Afterwards, a standardized dataset was constructed through data screening and normalization, combined with expert experience to manually annotate the bald patches in the images. Then, based on deep learning techniques, four classical network frameworks, Hrnet, Deeplabv3+, PSPnet and Unet, were built and paired with different backbone networks for model training respectively. The final results showed that Hrnet had the best recognition results, with the highest mean values of 68.27%, 75.90%, 78.22% and 99.22% for the Mean Intersection over Union, Mean Pixel Accuracy, Recall and Accuracy, respectively. In summary, the results showed that it is effective to combine low-altitude UAV remote sensing platform with deep learning technology. This study provides a new method for the identification of alpine meadow bald patches.

Keywords

• deep learning
• segmentation
• remote sensing
• degraded alpine meadows
• patches