Roof damage detection and evaluation using aerial image based on improved DeepLabv3+

To address the inefficiency and low accuracy of traditional roof damage detection methods, this study proposes an improved DeepLabV3+ model, which integrates the Swish activation function, Efficient Channel Attention (ECA) mechanism, depthwise separable convolution (DSConv), and an optimised ASPP module ECA-ASPP, combined with a refined MobileNetV2 backbone network. Compared with transformer-based architectures, the proposed model achieves a better balance between computational efficiency and segmentation accuracy, making it more suitable for real-time UAV deployment. Additionally, the model demonstrates strong generalisability in cross-scene validation and maintains stable performance under limited training data conditions, to achieve high-precision roof damage detection in aerial images. For validation, a dataset comprising 7,108 multi-scene roof images was constructed. Experimental results demonstrate that the proposed model achieves the mIoU of 85.86%, the mPA of 91.64%, and a processing speed of approximately 35 FPS, outperforming existing state-of-the-art methods. This research provides theoretical and technical support for UAV-based intelligent inspection systems, contributing to advancements in automated roof damage detection.