AgroConform: Coverage-Guaranteed Crop Recommendation with Sensor-Noise-Aware Calibration, Distribution-Shift Monitoring and Cost-Aware Agronomic Recourse

Abstract

Crop recommendation models now exceed 99% accuracy on standard soil and climate data, yet a single predicted crop does not show how far that advice can be trusted in a given field. This study presents AgroConform, which wraps any probabilistic classifier with split conformal prediction and maps each output to one of three actions: recommend, shortlist or abstain. On 2,200 records covering 22 crops, the framework met its 95% coverage target, and recommended crops were correct in 99.6% of cases. Noise-aware calibration held coverage at 95% under simulated soil-sensor error, whereas standard calibration fell to 66%. A shift monitor flagged 49% of fields after a simulated 6 °C warming, while the classifier changed its advice for fewer than 1%. A cost-aware recourse search priced the soil amendments that make a preferred crop admissible and showed that most refusals are climatic. Each field takes about 15 ms on a standard CPU.