CRDS-Unet: channel-reversed depthwise separable U-net for efficient tomato leaf disease segmentation

Introduction Automated plant leaf disease segmentation is crucial for accurate disease diagnosis and quantitative lesion assessment in precision agriculture. U-Net-based architectures are widely employed for this task, although their high computational complexity sometimes hinders their application in resource-constrained agricultural settings. Methods This work investigates depthwise separable convolutions in conjunction with a channel-reversed U-Net for tomato leaf disease segmentation. The channel-reversed encoder allocates more feature channels to the shallow layers and progressively fewer to the deeper layers, the reverse of a conventional U-Net’s channel schedule, with the aim of preserving spatial detail while reducing computation in the deeper, more computationally expensive stages. Depthwise separable convolutions are used throughout the encoder, bottleneck, and decoder to further reduce trainable parameters. The model was trained on lesion masks and evaluated on an independent, held-out subset of the PlantVillage tomato leaf dataset using Dice coefficient, IoU, precision, recall, and F1-score, following a corrected data-splitting protocol that eliminates train/test leakage present in an earlier version of this work. Results The proposed method reduced trainable parameters by over 97% relative to a conventional U-Net (0.96 million versus 31.06 million), at a measured reduction in segmentation accuracy under the training budget evaluated (Dice coefficient 0.79 versus 0.90 for standard U-Net). A supplementary experiment with a longer training budget narrowed this gap substantially (Dice coefficient 0.85). Discussion These findings indicate that the parameter-efficient design requires more training to approach the accuracy of higher-parameter alternatives. We report this accuracy/efficiency trade-off directly, rather than claiming accuracy parity, and identify a full matched-budget comparison across all evaluated architectures as a priority for follow-up work. By demonstrating a substantial, disclosed parameter-efficiency gain, this approach is intended to support future development of low-cost disease diagnosis tools for resource-constrained agricultural settings. Validated real-time, on-device deployment remains a direction for future work.