A new machine learning mango damage segmentation dataset and application models

Mango fruit diseases significantly impact both market quality and yield. Early and accurate detection is of paramount importance for supporting farm and processing decisions. However, traditional manual inspection is typically slow, subjective, and error-prone. This research proposes a deep learning-based approach for segmenting diseased regions on mango fruits by comparing two popular semantic seg-mentation models, U-Net and DeepLabV3. A dataset of 1000 mango fruit images was acquired under natural lighting conditions, pixel-level, and augmentation was carried out which gave 4167 images, and after all the preprocessing steps, grayscale K-means masks were generated to provide ground truth. The two models were then trained and evaluated using a total of 500 images and 500 ground truth masks. Experimental results indicated that DeepLabV3+ showed better segmentation performances and lesion boundary estimation with stronger generalization compared to U-Net. These results confirm the ability of deep learning-based segmentation to provide reliable pixel-level lesion localization that can be used for automatic and objective assessment of mango disease for practical applications in agriculture.