Mapping of Rodent-Induced Disturbance Impact on Biological Soil Crusts via Deep Learning-Informed Adaptive Graph Cut

Biological soil crusts (biocrusts) function as a "living skin" at the soil–atmosphere interface, covering approximately 12% of the global terrestrial surface and providing critical ecosystem services. While climate change and anthropogenic pressures are recognized drivers of biocrust degradation, the extensive disturbance induced by rodent communities represents a pervasive yet under-recognized threat. Rodent burrowing activities mechanically destroy biocrusts, creating a distinctive landscape mosaic of high-albedo, excavated soil patches set against the darker, intact biocrust matrix. In this study, we present a hybrid mapping framework that integrates both data- and model-driven approaches to generate the first regional-scale, multi-year (2017–2025) spatiotemporal map of rodent-induced disturbance in the Gurbantunggut Desert, China. Our methodology employs a two-stage strategy: (1) To identify potential disturbance areas, a Swin Transformer segmentation model is trained using semi-automatically generated pseudo-labels leveraging thresholding based on the brightness contrast between burrows and biocrusts. (2) Final boundaries are then optimized through an adaptive Graph Cut algorithm that integrates deep-learning probability maps with morphological priors and spatial gradient information. Validated against 125 field-surveyed sites, the framework achieved an overall accuracy of 0.95, with specific F1-scores for rodent-induced disturbance and background reaching 0.81 and 0.98, respectively. Our analysis revealed that rodent-induced disturbances followed a "rise-and-fall" temporal trend, peaking around 2019. At its peak, the disturbed area accounted for 8% of the entire desert region, representing a striking 23% of the total biocrust coverage. This work offers a reliable methodology and dataset to assess and understand the neglected role of bioturbation in dryland ecology. Our study is highly relevant for dryland conservation, exemplifying how bioturbation shapes desert ecosystem stability and the functional integrity of biocrust-dominated landscapes.