Cross-Scale Mixture-of-Experts with Conformal Uncertainty for Reliable Crop Yield Prediction across Latent Agro-Climate Regimes

Abstract

Accurate and reliable crop-yield prediction is foundational for climate-resilient agricultural planning, yet data-driven models are commonly developed either for micro plot-scale records or macro country-level panels in isolation and often without decision-ready uncertainty. This paper presents AgroX-MoE-CP, a cross-scale mixture-of-experts (MoE) framework equipped with split conformal prediction to deliver calibrated prediction intervals across heterogeneous agro-climatic regimes. First, latent agro-climate regimes (AgroClusters) are discovered via unsupervised clustering of standardized weather and management variables, capturing heterogeneity without manual zoning. AgroClusters are then provided as an additional categorical context to a tabular MoE model, where a gating network adaptively blends expert predictors. On a synthetic plot-level dataset (1M records), we inject agronomy-aware engineered features (rainfall intensity and an approximate growing degree days indicator). On a FAO-based macro panel, we stabilize learning by regressing on a log-transformed yield and map predictions back to the original units. Empirically, AgroX-MoE achieves strong point accuracy at both scales (micro: R²≈0.91, MAE≈0.40 t/ha; macro: R²≈0.88, MAE≈1.68 t/ha equivalent). Split conformal calibration yields finite-sample-valid 90% prediction intervals with empirical coverage ≈0.90 on both datasets and moderate average widths (micro: ≈1.66 t/ha; macro: ≈8.88 t/ha equivalent). Results indicate that combining latent regime discovery, MoE specialization, agronomic feature engineering, and distribution-free conformal calibration supports accurate and reliable crop-yield forecasting across scales.