AtroUNet: Atrous Convolution & ResNet-34 Based U-Net Architecture for Land Cover Classification

Satellite image segmentation faces challenges due to spatial and spectral variations, as well as unclear object boundaries caused by the limitations of regular convolution and atmospheric factors. This study applies a U-Net architecture with Atrous Spatial Pyramid Pooling (ASPP) and image enhancement techniques to improve land cover segmentation accuracy. Training results indicate that using a pretrained ResNet-34 accelerates model convergence, stabilizing metrics after 30–40 epochs. The AtroUNet model with rotating augmentation achieved the best performance, with 79.83% accuracy and a mean Intersection over Union (mIoU) of 0.77. However, contrast stretching and edge enhancement did not consistently improve performance, highlighting the need for selective image enhancement techniques. This study confirms that ASPP enhances land cover classification performance, though further optimization in image preprocessing and model architecture is necessary.