Medium-resolution land cover products are the main data sources for large-scale ecological and environmental research. However, medium-resolution remote sensing imagery lacks spectral detail and texture prominence, especially in complex and fragmented terrain scenes, where classification accuracy is low. To response to this issue, this paper constructs a Double-Attention U-Net model that integrates a spatial attention mechanism module at the initial stage of the encoder to enhance the capture of texture features. In the fourth stage, a channel attention mechanism is introduced to enrich spectral features through cross-layer information fusion, thereby enhancing the overall model’s ability to depict local details. The ablation experiment verifies that the mF1-score is improved by 1.31% after adding the spatial attention mechanism and channel attention mechanism modules. Based on this model, this paper uses GF-1 WFV image to obtain the 2016 and 2020 land cover classification maps of Nepal. Through the combination of overseas experimental data acquisition and internal processing, a set of high-precision and widely distributed verification dataset is formed. The results of the accuracy validation study based on this data set show that the average overall accuracy is 83.33%, and the classification effect performs well in the fragmented feature scenarios of Nepal’s urban and peripheral areas, in which the overall accuracy is 82.79% in 2016 and 83.87% in 2020.