<![CDATA[The selection of a convolutional neural network (CNN) architecture for cocoa (Theobroma cacao) pod disease detection involves a trade off between classification accuracy and computational efficiency that is decisive for eventual deployment on the mobile hardware available to smallholder farmers. This study presents a controlled comparative simulation of three widely used architectures, EfficientNetB0, ResNet50, and MobileNetV2, under identical, literature-grounded conditions. Rather than reporting field-validated results, a balanced synthetic dataset of 3,000 images spanning four classes (healthy, black pod, pod borer, frosty pod) was generated with class-conditional feature statistics parameterized from published references. All three models were initialized with ImageNet weights, fine-tuned with an identical training protocol and shared data splits, and evaluated on the same held-out test set. In simulation, EfficientNetB0 achieved the highest accuracy (93.8%) and macro F1 (0.938), followed by ResNet50 (92.7%, 0.926) and MobileNetV2 (91.1%, 0.909). When efficiency is considered, the ranking shifts: MobileNetV2 offered the smallest footprint and lowest latency, EfficientNetB0 delivered the best accuracy-per-parameter, and ResNet50 was the most resource-intensive without a commensurate accuracy gain. The dominant error mode across all models was confusion between pod borer and frosty pod. The results indicate that EfficientNetB0 offers the most favorable accuracy efficiency balance for this task, while MobileNetV2 is preferable under strict on-device constraints. All figures are framed explicitly as simulation outputs and discussed in light of the synthetic-to-real domain gap]]>
Copyrights © 2025
