Maturity Grading of Cashew Apple Fruit Using Deep CNN Models

Cashew apple fruits play a vital role in the coastal economies of regions such as Goa, India, where they are cultivated widely for fresh fruit consumption and juice production. This fruit is shipped worldwide, and its market value is closely tied to its quality, which is affected by its stage of ripeness. This paper introduces an automated system for grading the maturity of cashew apples using deep convolutional neural networks (CNNs). Several pretrained models – AlexNet, ResNet-50, MobileNetV2, VGG19, and custom-designed CNN—were employed to classify cashew apples into three categories: overripe, ripe, and unripe. A dataset comprising images from two different farms in Goa, India, was pre-processed, and later augmented to improve training and validation. The custom CNN achieved the highest classification accuracy of 78% with superior F1-score (0.83) for overripe and unripe grades. It comprises model parameters (17.2 M), model size of 68.8 MB, and real-time inference (191 FPS, 5.24 ms). MobileNetV2, however, achieved superior performance for the ripe grade, with an F1-score (0.71) and remained the most lightweight architecture. The result was further validated using a 5-fold cross-validation and integrated gradients which confirmed reliable feature attribution. This study offers a promising foundation for automated cashew apple grading and future drone-based deployment.