AeroUNet3+: An Attention-aware UNet3+ Architecture for Enhanced Semantic Segmentation of UAV Aerial Imagery

This paper proposes AeroUNet3+, an enhanced UNet3+ model incorporating the Convolutional Block Attention Module (CBAM) and Atrous Convolution to improve segmentation accuracy in complex UAV imagery. CBAM refines feature selection using channel and spatial attention, while Atrous Convolution expands the receptive field without added computational cost. We incorporate Focal Loss and data augmentation to mitigate class imbalance. Extensive experiments on the AeroScapes and UAVid datasets demonstrate improved generalization. Notably, our proposed model achieves an mIoU of 67.58% on AeroScapes and 49.60% on UAVid, surpassing baseline UNet3+ and other recent Deep Learning models. The results confirm that the introduction of CBAM and Atrous Convolution provides a lightweight yet effective solution for UAV-based segmentation, balancing computational efficiency with high segmentation accuracy.