Hybrid Sensor-Vision Machine Learning for Predicting Mahewu Fermentation Dynamics.

This study developed and evaluated a laboratory-scale hybrid sensor-vision machine learning framework for monitoring and predicting mahewu fermentation dynamics using synchronized physicochemical sensing and surface-image analysis. White- and yellow-maize mahewu were fermented under controlled processing conditions while pH, electrical conductivity (EC), and temperature were recorded and surface images were captured under fixed LED illumination. Images were segmented and converted into color, texture, edge, and morphology descriptors, then aligned with sensor records and raw laboratory total titratable acidity (TTA) and total soluble solids (TSS) measurements using tolerance-valid matching. Semantic segmentation was evaluated separately across 96 manually annotated images from 14 fermentation-run groups using nested grouped validation, producing a run-balanced mean Dice coefficient of 0.8070 and mean intersection over union of 0.7195. Primary raw laboratory prediction models used 7945 matched image-sensor records from 12 main fermentation runs, including 71 TTA and 87 TSS reference records. Leakage-controlled models were evaluated by fermentation run using image-only, sensor-context, and multimodal feature sets. Raw laboratory TTA was predicted most effectively using sensor-context variables with Gradient Boosting (R