Lightweight Pixel-Wise Lesion Segmentation and Severity Estimation of Tobacco Leaf Disease from Close-Up Images: A Group-Aware Benchmark with Robustness Analysis

Lesion area is the basis of disease severity rating in tobacco, but visual rating is subjective, and most deep-learning studies classify whole images on random splits that can leak near-duplicate photographs. This study evaluates lesion segmentation and severity estimation on 3,485 close-up tobacco leaf images (800 × 800 px) with pixel-level masks, using group-aware training, validation and test partitions (3 independent splits) and a lightweight pixel-wise gradient-boosting classifier operating on colour, texture and image-normalised features. The full model reached a Dice coefficient of 0.837 ± 0.020 (IoU 0.720 ± 0.029), compared with 0.187 for the best Otsu-threshold baseline. Predicted lesion area agreed closely with annotation (Pearson r = 0.981; mean absolute error 0.39 percentage points) and reproduced a four-grade severity scale with 90.0% accuracy (quadratic-weighted κ = 0.913). Without retraining, texture features improved robustness to occlusion and brightness change but caused a collapse under Gaussian noise (Dice 0.143); noise-aware training and a median pre-filter restored Dice to 0.763 and 0.666, respectively. Inference took about 79 ms per image on one CPU core. The data carry no disease-class labels, so no pathogen identification is claimed. The results provide a reproducible CPU-only baseline and a transparent account of failure modes for lesion-level phenotyping of tobacco.