CDUnet: a plant leaf segmentation model integrating spatial coordinate attention and dynamic upsampling

High-precision plant leaf segmentation is a critical component for automated analysis in precision agriculture and plant phenomics, yet it faces challenges in scenarios with complex backgrounds, overlapping leaf occlusions, and small-scale leaves. To address these issues, this paper proposes CDUnet, an improved U-Net model. The model employs the VGG16 backbone network in the encoder to enhance feature extraction. A Coordinate Attention (CA) module is embedded in the skip connections, fusing spatial coordinate information to strengthen the localization of key leaf structures and suppress background interference. In the decoder, a Dynamic Upsampling module (Dysample) replaces traditional interpolation, enabling adaptive detail reconstruction to improve segmentation accuracy for leaf edges and small leaves. Additionally, a hybrid loss function combining Dice Loss and Focal Loss is adopted to alleviate the class imbalance problem. Experiments demonstrate that the proposed method significantly outperforms baseline approaches in precision, F1-score, Intersection over Union (IoU), and mean Intersection over Union (mIoU), providing an effective solution for high-precision segmentation of small leaves against complex backgrounds.