Land-Cover Classification and Change Analysis in an Arid Oasis Based on Deep Learning and High-Resolution Remote Sensing Imagery: A Case Study of Minqin County, Gansu Province

Reliable land-cover classification in arid oasis regions is imperative for pivotal decisions concerning oasis stability assessment, desertification control, and the optimal allocation of water resources. However, such regions are characterized by fragmented ground objects, inter-class homogeneity, intra-class heterogeneity, and multiscale spatial distributions, making traditional classification methods based on medium- and low-resolution remote-sensing imagery inadequate for fine mapping. To this end, this study proposes a high-resolution U-shaped Mamba network (HRUMamba). By integrating long- and short-range contextual modeling, multiscale feature extraction, and attention mechanisms, the model enhances the representation and learning capabilities for complex land-cover features in arid oasis regions. Specifically, HRNet is employed as the encoder to capture local detail features and model short-range contextual relationships, whereas the improved scale visual state space (SVSS) block serves as the core of the decoder to model long-range contextual relationships and extract global semantic features. In addition, the adaptive awareness fusion module (AAFM) embedded in the skip connections can effectively suppress irrelevant noise interference and improve the model’s feature learning capability for small-scale ground objects and minority-class samples. Experimental results show that the model achieves optimal performance on the Minqin-10 dataset, with mF1 reaching 90.56%. The land-cover classification results of Minqin County based on Gaofen-2 images show that the overall accuracies in 2018 and 2023 are 90.80% and 91.12%, respectively, and the Kappa coefficients are 0.8746 and 0.8835, respectively. The two-period land-cover change analysis between 2018 and 2023 showed that the areas of farmland and bare-land decreased by 107.19 km2 and 51.45 km2, respectively. Farmland was mainly converted to grassland and bare-land, whereas bare-land was mainly converted to grassland and woodland. The annual change areas for farmland and bare-land are –17.87 km2 and –8.58 km2, with annual change rates of –1.43% and –0.06%, but their dynamic degrees remain high, indicating active land-cover change. The net change areas of other land-cover types all increased. This study further verifies the effectiveness and application potential of HRUMamba for land-cover classification of high-resolution images in arid oasis regions.