IoT-Based Monitoring and Assisted Classification of Cocoa Bean Fermentation Using Multimodal Sensor Data

Cocoa fermentation is commonly evaluated through periodic observation, although temperature, humidity, pH, and bean appearance change continuously and may overlap between stages. This study evaluates an IoT-based monitoring and assisted-classification system for under-fermented, optimal, and over-fermented conditions. Internal model development used 716 sensor records from three source batches and 2,154 labeled images. Sensor-only Random Forest was evaluated using leave-one-batch-out cross-validation, ResNet50 used group-aware image partitions and transfer learning, and hybrid Random Forest combined 17 visuals with eight sensor features. Internal accuracies were 98.60%, 94.12%, and 98.70%, respectively. To remove deployment-batch leakage, 216 images identical to field-batch images were excluded, and the models were refitted using 1,938 non-field images. Three physically implausible sensor records were also excluded. On the resulting 224- record field holdout, sensor-only and hybrid accuracies were 97.32% and 96.88%; their difference was not significant. Sensor-derived variables contributed 78.85% of hybrid feature importance, explaining the absence of a consistent fusion advantage. The system supports continuous logging and assisted stage classification but does not demonstrate autonomous process control or failure prevention. Because field labels were operational rather than laboratory-confirmed, further multi-batch evaluation using expert or biochemical ground truth remains necessary.