DOMAIN ADAPTIVE VERMICOMPOST QUALITY PREDICTION THROUGH CALIBRATED MULTI-MODAL LEARNING AND INTERNET OF THINGS

Abstract

ABSTRACT

Vermicomposting recycles organic wastes into nutrient-rich amendments that will have a market value and agronomic performance that depends upon the consistency of quality. Producers require quick, objective measures of maturity and safety in order to schedule harvests and certify batches. However, field labs are expensive and slow, single-indicator proxies (e.g., temperature decay) are not reliable across feedstocks, and most ML studies involve poor datasets with poor cross-site generalization. There is a gap for a low-cost, multi-modal, and transferable predictor that works in actual plants. To overcome such a gap, in this study, VERMI-Quality Network (VERMIQualNet) is suggested, which is a multi-modal and domain-adaptive real-time learner that integrates Internet of Things (IoT) with low costs, Near Infrared/Red-Green-Blue (NIR/RGB) spectrum, and texture/color features based on the bin images using self-supervised encoders and the gradient boosted multi-task head. The model is calibrated using uncertainty-aware temperature scaling and batch-wise transfer to output (i) Maturity Index, (ii) Contamination Risk, (iii) Nutrient Score, and in addition, Feedstock-Aware Pattern Mining (FAPM) is also integrated and extracts cross-batch association rules, which identifies hidden co-evolution trends between maturity and contamination. Interpretability is given through SHAP to reveal drivers (e.g., humification signal). Methodological core processes include cross-feedstock, data acquisition, cross-season, signal cleaning, and feature learning by contrastive pretraining. The processes are then ultimately deployed as an edge deployable inference service.

Keywords:- vermicompost, transparency, mining, machine learning, training, scaling, bin, images, features, multi-modal, IoT