CIA-UNet: An Attention-Enhanced Multi-Scale U-Net for Single Tree Crown Segmentation

Accurate segmentation of single tree crowns enables precise measurement of tree height, DBH, and stock volume, facilitating effective assessment of forest growth and productivity. However, single tree crown segmentation currently faces several challenges, including insufficient segmentation accuracy, mis-segmentation and crown adhesion. To address these issues, this study conducts experiments using UAV data and proposes a model named CIA-UNet. In this model, the multiscale feature extraction capability of the InceptionV2 module is leveraged to compensate for U-Net's limitation in capturing different scales of tree crowns and fine details using fixed convolutional kernels. The Convolutional Block Attention Module (CBAM) is employed to increase the weight of key features, and mitigate false segmentation caused by feature confusion. Additionally, the Attention Gate module is utilized to improve the model's sensitivity to foreground pixels, adjust the weighting of skip connection features and prevent segmentation errors due to semantic mismatches. The experimental results show that CIA-UNet achieves outstanding performance on TREE512 dataset, with overall accuracy (OA), mean precision (mPrecision), mean recall (mRecall) and mean intersection over union (mIoU) reaching 90.79%, 88.08%, 89.41% and 79.9% respectively. Compared with U-Net, these metrics have improved by 1.07%, 1.63%, 1.5%, and 2.41% respectively. Furthermore, CIA-UNet performs better than DeepLabV3+, PSPNet and FCN in accuracy evaluation indicators and visual segmentation results. The results demonstrate the efficacy of CIA-UNet model in enhancing the accuracy of tree crown segmentation and minimizing mis-segmentation, and provide experience for the study of single tree crown segmentation in the Millennium Forest of Xiongan New Area.