Answers are grounded in this paper and cite their sources.
Machine-learning crop recommenders are increasingly proposed for climate adaptation, yet the benchmark data behind them are rarely examined. We audited the public Crop Recommendation Dataset (2,200 records, 22 crops, seven soil and climate inputs), compared eight classifiers in a leakage-controlled pipeline and submitted the fitted models to idealized perturbations of temperature (+1.5 to +4 °C) and rainfall (±10 and ±20%). Each crop occupies fixed-width, uniformly filled ranges, a structure consistent with records generated from predefined intervals. Random Forest reached 99.36% five-fold accuracy, statistically indistinguishable from Extra Trees, while label permutations stayed at chance; thirteen engineered features did not raise accuracy. Rainfall changes reassigned up to 8.0% of records, against at most 0.6% for warming alone; the hottest and driest case reassigned 10.1% and placed 17.2% of records outside the applicability domain. Other classifiers responded more strongly but ranked the perturbations alike. A generic alert flag rose from 15.7% to 18.3%, a value governed by crop composition and threshold choice. The predicted class reproduces benchmark labels, not validated agronomic suitability; the protocol offers a transparent stress test for crop classifiers before they inform climate adaptation.
No takes yet. Share an insight, caveat, or question.
Al-Azzawi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: