AgriFuse-SSL: Region-Adaptive Weak-Strong Filtering and Confi-dence-Calibrated Semi-Supervised Learning for IoT Crop-Yield De-cision Support

Abstract

Agricultural Internet-of-Things deployments produce large, heterogeneous streams whose labels are sparse, whose sensor quality varies across seasons and regions, and whose class distributions can drift with crop and management conditions. These properties limit the reliability of crop-yield decision systems trained either on a comparatively small labeled regional dataset or on a larger but weakly supervised sensing pool. This paper proposes AgriFuse-SSL, a unified region-adaptive framework that combines the complementary mechanisms of two existing project models: the weak-strong filtering, informative-sample selection, and reliability-weighted ensemble model; and the pseudo-label generation, confidence-based selection, expanded sensing schema, and optimized semi-supervised convolutional predictor of large scale. The proposed framework first locks a 53-predictor schema, isolates independent farm-field records by record hash and time-region group, and applies weak and strong reliability gates that jointly consider missingness, local density, physical-range violations, temporal inconsistency, and model disagreement. It then forms density-calibrated pseudo-labels, corrects their regional prior, selects features using cross-region stability, and learns a compact feature-token encoder that combines one-dimensional convolution with tabular self-attention. A reliability-weighted regional ensemble produces low, medium, or high yield classes together with calibrated confidence and an accept-or-refer decision.  The source specification contains 1,000,000 Ensemble model records with 35 predictors and 3,000,000 Scalable model records with an expanded 53-predictor schema.  On a 100,000-record simulated test set, AgriFuse-SSL achieved 0.990 accuracy, 0.990 macro-precision, 0.990 macro-recall, 0.990 macro-F1, and 0.996 macro-AUC. The mechanism ablation increased accuracy from 0.971 for the supervised anchor to 0.990 for the complete model, while the proposed compact implementation reduced simulated batch latency to 72 ms. Region-stratified, calibration, risk-coverage, feature-stability, and uncertainty analyses show that proposed model more effective than treating filtering, pseudo-labeling, and prediction as independent stages.