Dual-Stage Deep Learning Framework for Hybrid Coconut Maturity Classification and Harvest Timeline Prediction

Coconut maturity estimation is critical in agriculture, as harvesting directly affects product quality, oil yield, and economic returns. Traditional methods, such as tapping or visual inspection, are subjective and inconsistent. This research develops a dual-stage deep learning framework that enables classification of coconut maturity stages and prediction of harvest timelines. The framework integrates a hybrid convolutional neural network (EfficientNetB0 + DenseNet121) for classification with a MobileNetV2-based regression model for predicting harvest time in immature coconuts. Images were collected, preprocessed, and augmented to balance classes. The models were trained and validated using accuracy, F1-score, mean absolute error, and root mean square error. A Gradio-based web application was developed to enable real-time image upload, classification, and timeline estimation. The hybrid classifier achieved over 99% accuracy, outperforming single-model baselines, while the regression model recorded an MAE of 36 days and an RMSE of 27 days, confirming reliable predictions. The web interface demonstrates practical usability and accessibility for farmers. While the dataset was limited in size and scope, which may affect generalizability, this study introduces the first dual-stage coconut framework that combines classification and predictive modeling into a practical, scalable system deployable on mobile and edge devices. Beyond its practical contributions, the study also advances agricultural AI research by extending coconut maturity studies from static classification into predictive modeling, a direction that remains underexplored. Future research will focus on expanding the dataset to diverse environments, integrating multimodal variables such as weather and soil data, and enhancing robustness under real-world conditions.