Classification of Avocado Varieties and Maturity Levels Using Quantized MobileNetV2 in a Mobile Application

Abstract

Visual grading of avocado ripeness and variety by distributors and consumers is subjective, and published deep-learning studies on avocado rarely report what happens when a trained model is integrated into a working application. This study presents a proof-of-concept, end-to-end system that classifies two Indonesian avocado varieties (Aligator and Miki) and four ripeness stages (unripe, half-ripe, ripe, overripe) from smartphone photographs. Two MobileNetV2 transfer-learning models were trained on a self-collected image set photographed from several fruits over multiple observation days, with training, validation and test subsets separated by fruit identity to prevent data leakage. The models were converted to post-training-quantized TensorFlow Lite (≈2.7 MB each) and served through a FastAPI–MySQL backend to a Flutter application. On a held-out test set of 32 images, the variety model classified all images correctly (100%; Wilson 95% CI 89.3–100%), and the ripeness model reached 84.4% accuracy (95% CI 68.2–93.1%), a macro-F1 of 0.814, a Cohen's κ of 0.686 and a quadratic-weighted κ of 0.900, against a majority-class baseline of 62.5% (one-sided exact binomial p = 0.006). Four of the five ripeness errors were half-ripe fruits predicted as unripe; half-ripe recall was 0.375 (3/8). Integration testing exposed two defects invisible to offline evaluation, a class-label ordering error and double input normalization, which caused near-constant predictions until corrected. Because the test set is small and drawn from a limited number of fruits, the results should be read as preliminary, and larger fruit-level evaluation and field testing with users are needed.