LARU-net: lesion-aware residual attention U-net for binary cucumber leaf lesion segmentation and continuous severity estimation
Abstract
Plant leaf diseases threaten agricultural productivity and motivate accurate automated monitoring. This study presents LARU-Net, a unified architecture for binary cucumber lesion segmentation (lesion versus background) with an auxiliary continuous severity-regression output on a bounded 0–4 scale. The network integrates a Disease Feature Enhancement Autoencoder, Residual Multi-Scale Encoder, transformer bottleneck, Attention-Guided Decoder, and Boundary Refinement Module. On the declared held-out test partition (1,080 images), the archived full-model results were Dice 0.9230, IoU 0.8660, Boundary F1 0.9350, and precision 0.9460; the previously emphasized Dice value of 0.9456 is a validation value and is now identified as such. Severity estimates yielded an overall MAE of 0.29, RMSE of 0.37, and R 2 of 0.90. These values are single-run point estimates from one greenhouse image collection. Plant/session identifiers, complete acquisition and annotation provenance, repeated-run logs, and an external dataset were not preserved in the supplied study archive; therefore, the results support in-domain evaluation but not broad cross-greenhouse or cross-season generalization. The split was stratified at image level; independence at plant, leaf, or acquisition-session level could not be verified. Confidence intervals and significance tests were not reconstructed because per-image predictions and repeated-run logs were unavailable.