A Hybrid Deep–Texture Feature Fusion Framework for Coconut Maturity Classification Using MobileNetV2 and Local Binary Patterns

Objectives: The main objective is to explore and devise a hybrid deep-handcrafted texture fusion approach for automated coconut image-based maturity classification using MobileNetV2 and Local Binary Patterns. This approach aims to enhance discrimination of three coconut maturity stages: tender, mature and overmature. Method: A newly created coconut dataset of 3050 coconuts annotated as tender, mature and overmature were used. The proposed framework integrates the lightweight MobileNetV2 CNN Model with handcrafted texture descriptor via feature fusion analogy. ImageNet is employed to pretrain MobileNetV2 model to extract deep features, while uniform LBP captures local texture information from coconut images. The feature vectors thus obtained are fused and prediction process is performed using Softmax classifier. The efficacy of the model was evaluated using three independent runs and standard classification metrics like accuracy, precision, recall and F1-score. Findings: The proposed MobileNetV2–LBP framework achieved 88.82% ± 1.67% mean validation accuracy across three runs, compared with 73.01% ± 0.19% for the MobileNetV2 baseline. The highest individual-run accuracy was 91.04%. A paired t-test across the three runs indicated a statistically significant improvement (t (2) = 14.75, p = 0.0046). The final run achieved 91.04% accuracy, with recalls of 97%, 86% and 86% and F1-scores of 0.95, 0.87 and 0.89 for tender, mature and overmature coconuts respectively. The results demonstrate that LBP provides complementary local texture information to MobileNetV2’s deep semantic characterization. The empirical results indicate effective deep–handcrafted feature fusion for three-stage coconut maturity classification when compared with CNN baseline. Novelty: The study evaluates an application-oriented hybrid model by fusing MobileNetV2 deep representations with LBP texture features which substantially improves three-stage coconut maturity discrimination by providing complementary semantic and local texture information. Keywords: Coconut maturity classification, MobileNetV2, Local Binary Pattern, Feature fusion, Smart agriculture