MD-HRNet: A Land use Classification Method Based on Unbalanced Samples

Land use/land cover classification (LULC) can reveal changes in land use patterns and assess the impact of human activities on ecosystems. This study utilizes Sentinel-2 high-resolution satellite images to classify land use in Haikou City. Due to the complexity of the actual land classes in Haikou City, which leads to sample imbalance in the visually interpreted dataset, this paper proposes an MD-HRNet model to solve the sample imbalance problem and utilizes the combination of multi-scale convolution and frequency domain to extract more information. During the experiment, the land use categories are divided into nine categories, and the experimental results show that the model classification results are better, including 43.82% for mIOU, 53.57% for recall, 55.72% for F1, and 82.99% for accuracy.