Abstract:

The application of computer vision in precision agriculture has demonstrated considerable promise in automated plant disease detection. However, the effectiveness of such approaches is strongly dependent on the availability of high-quality, domain-specific datasets, particularly for deployment on resource-constrained edge devices. This paper introduces CocoaDetectDB, a publicly available image dataset developed for the detection of cocoa plant diseases under Tiny Machine Learning (TinyML) constraints. The dataset comprises images of healthy cocoa pods and three major cocoa diseases—Cocoa Black Pod Disease (CBD), Cocoa Swollen Shoot Virus Disease (CSSVD), and Frosty Pod Rot (FPR)—captured under real-world field conditions and supplemented with openly accessible public data. Images were curated, cleaned, and resized to a uniform resolution of 112 × 112 pixels to support low-memory and low-power inference. To validate the suitability of the dataset for automated disease classification, baseline experiments were conducted using MobileNetV2 and a lightweight quantized TensorFlow Lite model. Experimental results demonstrate classification accuracies of 99.13% and 93.75%, respectively, indicating that CocoaDetectDB contains sufficiently discriminative features for both conventional lightweight models and TinyML deployment. The dataset is intended to support future research in cocoa disease detection, edge AI, and resource-efficient agricultural monitoring systems.