A grazing pressure mapping method for large-scale, complex surface scenarios: integrating deep learning and spatio-temporal characteristic of remote sensing

Grazing pressure is an effective indicator for measuring the intensity of grazing activity. However, existing studies focused on small-scale sample measurements, which cannot be directly applied to large-scale mapping and livestock management work. We proposed for the first time a novel large-scale grazing pressure mapping framework suitable for complex surface environments, making full use of multi-source and multi-temporal remote sensing data to address the issue of being unable to monitor grazing intensity and grassland conditions from a large-scale perspective. The primary contributions of this research are as follows. Firstly, for non-typical objects such as potential grazing areas, we proposed an efficient sample dataset construction method involving two steps: manual pixel samples providing spectral information and machine learning samples offering contextual information. This approach successfully applied deep learning semantic segmentation methods to the field of non-typical object recognition. The accuracy of extracting potential grazing areas in the Three-River-Source National Park reached 86.62 %. Additionally, the model’s generalizability was validated using two testing sites on the Qinghai-Tibetan Plateau. Secondly, based on the time series patterns of multiple remote sensing indices reflecting high-intensity grazing, we established a quantitative model for grazing pressure for the first time. This model for grazing pressure successfully eliminated the impact of grassland growth conditions on experimental results and effectively reflected intensive grazing behaviors in seven indicators across three remote sensing indices, applicable to large-scale regions with multiple ecosystems coexisting. Overall, our approach achieved, for the first-time, rapid mapping of grazing pressure directly from remote sensing images in complex surface scenarios. More broadly, this study demonstrated the significant role of remote sensing technology in identifying non-typical features and characterizing human activities.