Application of Improved SeMask-FPN Model in Plateau Pika Burrows Recognition

The Three Rivers Source Region, situated on the Qinghai-Tibet Plateau, is an important natural reserve in China, with its grassland ecosystem playing a crucial role in maintaining regional ecological balance and water conservation. However, in recent years, the area has experienced severe grassland degradation due to factors such as overgrazing, grassland fires, climate change, and rodent infestations. Therefore, the urgent need for grassland restoration and management has emerged. To effectively address grassland degradation, accurate identification of its severity levels (e.g., first-degree degradation, second-degree degradation, etc.) is essential, followed by the implementation of appropriate restoration measures corresponding to each level. This study focuses on rodent infestations as a perspective to investigate grassland degradation. Through surveys, it has been found that the coverage ratio of rodent burrows, specifically those made by plateau pikas and field mice, can be used to determine the degree of grassland degradation. Accordingly, this research attempts to utilize semantic segmentation techniques to identify rodent burrows, particularly those of plateau pikas, and subsequently calculate the coverage ratio, enabling the determination of grassland degradation levels. Currently, due to the absence of a comprehensive publicly available rodent burrow dataset, we have conducted on-site data collection in the Three Rivers Source Region to establish a dataset based on rodent burrows. Additionally, we have improved the semantic segmentation model, SeMask-FPN. Addressing the insufficiency of global interactions in the SeMask-FPN model, we propose two different global fusion models: the first incorporates burrows convolution and self-attention mechanisms, introducing the Atrous Self-Attention (ASA) module to create the ASA-based SeMask-FPN model. The second model combines pooling and self-attention mechanisms, introducing the Pooling Self-Attention (PSA) module to establish the PSA-based SeMask-FPN model. Experimental comparisons reveal that the ASA-based SeMask-FPN model and PSA-based SeMask-FPN model achieved mIou values of 86.43% and 81.9%, respectively, representing improvements of 11.4% and 6.87% over the original SeMask-FPN model's mIou value.