AI-Based Cocoa Pod Disease Detection to Support Farmers in Upper Denkyira

Cocoa farmers in Upper Denkyira East Municipality, Ghana, can lose trees and harvests when pod diseases are noticed too late, and manual field inspection is often the only diagnostic option available. This paper presents a lightweight, mobile-deployable image classification system for early cocoa pod disease awareness. We apply transfer learning with a MobileNetV2 convolutional neural network, pre-trained on ImageNet, to a public cocoa pod image dataset across three classes: healthy, black pod rot, and pod borer. The model is paired with a farmer-facing advisory layer that translates each prediction into short, actionable guidance. Based on the test-set confusion matrix (n = 448), the model achieves approximately 88.2% overall accuracy, with strong performance on the majority "healthy" class (recall ≈ 0.97) but a substantially weaker recall on black pod rot (≈ 0.60) and an extremely small, statistically unreliable pod borer test sample (n = 4). Training curves further show validation accuracy peaking at epoch 4 (90.7%) before declining at epoch 5, an early sign of overfitting that this paper discusses openly. We situate these results against the parallel cassava leaf disease work by the same authors, discuss the model's limitations — including its reliance on secondary, non-Ghana-collected imagery, severe class imbalance in the pod borer class, and the absence of early stopping — and outline a path toward field validation and offline mobile deployment via TensorFlow Lite. This work demonstrates that existing, publicly documented deep learning techniques can be adapted at low cost into a locally-relevant decision-support tool for cocoa farmers in Upper Denkyira East.