洞庭湖湿地长时序动态监测与生态稳定性分析

Wetlands, due to their unique hydrological characteristics, experience drastic land cover changes. The lack of long-term, high-frequency land cover data has made it challenging to dynamically monitor these changes in wetlands. This paper presents a dual-weight adaptive semantic segmentation model that combines feature change intensity and category balance, developed based on long-term temporal and spatially fused imagery. This model proposes a high-frequency land cover classification mapping method, which effectively addresses the issues of continuity, high spatiotemporal resolution imagery, scarcity of classification samples, and high dependency on manual intervention in land cover mapping. It enables dynamic monitoring of land cover changes on a monthly scale in the Dongting Lake wetlands from 2001 to 2020. The overall classification accuracy and Kappa coefficient of the proposed method are 86.78% and 0.76, respectively. Based on long-term, high-frequency land cover change data, the land cover distribution in the Dongting Lake wetlands is closely related to the water level at Chenglingji station. The ecological stability shows an annual characteristic of initially decreasing then increasing, along with a seasonal characteristic of being higher in summer and lower in winter. The results of this study provide empirical evidence for targeted wetland conservation decision-making and flood management.