A Reliability-Aware Distributionally Invariant and Uncertainty-Selective Framework for Multi-Regional Crop Yield Classification

Abstract

Agricultural Internet-of-Things data differ greatly between farms, regions, seasons, and sensor deployments. A predictor learned on pooled data can have high average accuracy but break when missingness, sensor drift, class imbalance or geographic covariate shift modifies the input distribution. In this paper, we introduce RADIUS-Yield, a novel extension for low, medium, and high crop-yield classification. The framework substitutes discrete hard filtering with continuous reliability weights, substitutes entropy-only sampling with a class-region constrained coreset, connects regional experts via a similarity graph, modifies global and local probabilities via a shift gate, and yields conformal prediction sets with an optional rejection option. Five algorithms follow the architecture sequence, and five propositions prove bounded reliability weights, monotone class-region coverage, convex graph regularization for fixed representations, probability simplex preservation, and marginal conformal coverage under exchangeability. Raw field records were not accessible for a new run. A seeded controlled experiment was simulated with 18,000 instances, 24 features, six regions, 1,800 farm groups, region-dependent missingness, sensor faults, nonlinear class structure, and an induced test shift. RADIUS-Yield achieved a macro F1 of 0.7500 and an AUROC of 0.9174 with an expected calibration error of 0.0219 and a farm-bootstrap 95% macro-F1 interval of [0.7361,0.7631]. Multinomial logistic regression resulted in the best macro F1 at 0.7583, so the experiment does not support the assertion of a method being superior in general. Under the enforced shift, the 90% conformal set attained 88.63% coverage, revealing the limit of exchangeability-based calibration. The results imply the internal functioning of the novel reliability, regional, and selective mechanisms and delineate the field-data experiments required prior to making deployment assertions.