<![CDATA[Automated detection of cocoa (Theobroma cacao) pod diseases such as black pod, pod borer infestation, and frosty pod rot is critical for safeguarding yield, yet the development of deep-learning classifiers is frequently constrained by the scarcity of curated, well-balanced image datasets. This study presents a controlled simulation that evaluates the expected performance envelope of an EfficientNetB0 classifier under idealized, literature-grounded conditions before field data collection is undertaken. Rather than asserting empirical field results, a synthetic dataset is constructed whose per-class feature distributions (color, texture, and lesion morphology) are parameterized from values reported across six core references. A balanced corpus of 3,000 synthetic images spanning four classes (healthy, black pod, pod borer, frosty pod) was generated and partitioned using a stratified 70/15/15 split. EfficientNetB0, initialized with ImageNet weights and fine-tuned with standard augmentation, achieved a simulated test accuracy of 93.8%, a macro-averaged F1-score of 0.926, and balanced per-class precision and recall in the 0.90-0.95 range. The confusion matrix indicates that the principal source of error is morphological overlap between pod borer and frosty pod presentations. The results delineate a plausible upper-bound performance band to guide sample-size planning, augmentation strategy, and architecture selection for a subsequent field study. All reported figures are framed explicitly as simulation outputs.]]>
Copyrights © 2025
