WaveUNet+: Preserving Root System Architecture Integrity in In Situ Root Segmentation via a Unified Spectral–Spatial Framework

Root phenotypic analysis is closely related to crop yield and stress resistance. Although deep learning can improve the efficiency of root phenotype recognition, existing methods suffer from insufficient segmentation accuracy under complex soil backgrounds and focus on a single target. To address the issues of limited accuracy and operational complexity in existing root segmentation models, this paper proposes a novel wavelet-enhanced full-scale segmentation network. The WaveUNet+ model is based on U-Net3plus, replaces traditional downsampling with the Haar wavelet transform, and introduces the EMA module. The impact of the wavelet transform is validated using Grad-CAM, and HD95 is employed to evaluate the improvement in segmentation quality brought by the attention mechanism from the perspective of boundary accuracy. Transfer learning is used to improve model generalization, and the test results on diverse roots and various soils are compared. A Docker containerized root image segmentation method is designed to achieve convenient and practical operation, and the deployment feasibility of the model on edge devices is also verified. Our model effectively enhances the recognition of fine roots in soil backgrounds, leading to improvements across various metrics, achieving an Accuracy of 99.2%, while improving model accuracy with relatively low parameter count and model size. Compared with the original U-Net model, mIoU is increased by 1.52% and Recall by 2.93%. The results show that the model not only performs excellently on the original dataset but also maintains good generalization ability across different imaging modalities, crop species, and soil conditions. With Docker, users can achieve root image segmentation on their own computers without tedious program installation and environment configuration. In the future, we will attempt methods such as pruning and quantization to reduce model size, so as to better adapt to the deployment requirements of edge devices.