Evaluation of U-Net Variants and Traditional Machine Learning Methods for Land Cover Classification Using High-Resolution Satellite Imagery

Abstract. Recent advancements in deep learning have led to improvements across fields including computer vision, biomedical engineering, and geospatial analysis. Deep neural networks (DNNs) are prior at extracting complex spatial features from large-scale data, enabling accurate automated interpretation of remote sensing imagery. Land cover classification and segmentation are critical for urban development, environmental monitoring, and agricultural planning. As high-resolution satellite data become more accessible, demand for precise classification methods grows. This study investigates DNN performance for land cover classification and segmentation using high-resolution DigitalGlobe satellite imagery with three bands, specifically the DeepGlobe Land Cover Classification dataset derived from WorldView-3 imagery. Six land cover classes are examined: urban land, agricultural land, rangeland, forest land, water, and barren land. The dataset is divided into 70% training, 20% validation, and 10% testing subsets. Deep learning models (AttUnet, Unet++, and U-Net) are implemented, with evaluation metrics including accuracy, F1-score, and recall. We analyze test data using trained models and compare results to traditional algorithms to assess robustness and generalization. Results show Unet++ achieved an F1-score of 84.21%, accuracy of 90.32%, and recall of 87.46%, demonstrating superior performance. AttUnet and U-Net followed with F1-scores of 82.12% and 77.08%. SVM and RF performed good but lower, with F1-scores of 65.64% and 70.21%. Unet++ showed better performance in classifying water and agriculture, excelling at identifying boundaries. This makes Unet++ the most effective model, especially in heterogeneous regions with complex landscapes. The comparative analysis highlights its advantages in achieving higher segmentation accuracy and spatial consistency.