CORINE Level-2 Land Cover Classification from Sentinel-2 Images Using U-Net Deep Learning Model

This study implements and evaluates an advanced U-Net deep learning architecture for semantic segmentation of multi-spectral Sentinel-2 imagery to classify CORINE (Coordination of Information on the Environment) Level 2 land cover classes in Türkiye. We implemented an advanced U-Net architecture incorporating spatial attention mechanisms and residual connections, utilizing multi-spectral Sentinel-2 imagery (Red, Green, Blue [RGB], Near-Infrared [NIR], and Short-Wave Infrared [SWIR] bands). A total of 92 working units were identified across Türkiye to represent all 15 classes of Level 2, covering 12,000 square kilometers (approximately 1.5% of the country's total area). The imagery was collected between June and September 2018, corresponding to the original CORINE 2018 production timeframe. The model was trained on 2,815 patches of 256x256 pixels generated from these working units. Our architecture integrates an Atrous Spatial Pyramid Pooling (ASPP) block for multi-scale feature extraction and employs a hybrid loss function combining Categorical Cross-Entropy and Jaccard Loss. The model achieved a mean Intersection over Union (IoU) of 0.5225 across all classes, with particularly strong performance in water bodies (Marine Waters: 0.8884 IoU, 0.9291 accuracy; Inland Waters: 0.8113 IoU, 0.8791 accuracy) and permanent crops (0.7118 IoU, 0.8389 accuracy). While the model excelled in classifying spectrally distinct classes, it faced challenges with classes sharing similar characteristics, such as artificial non-agricultural vegetated areas (0.0896 IoU, 0.1017 accuracy) and pastures (0.2549 IoU, 0.3504 accuracy). This research demonstrates the potential of deep learning approaches for automated CORINE land cover mapping, while highlighting areas for future improvement in distinguishing between spectrally similar classes.