Work

Benchmark Accuracy Is Not Field Reliability: Auditing Crop and Fertilizer Recommendation Datasets and Their Robustness to Sensor Noise

Year: 2026
Type: preprint
Abstract: Machine-learning crop and fertilizer recommenders are usually judged on clean public benchmarks, where accuracy above 95% is routinely reported. We test whether such scores reflect field reliability. On the Crop Recommendation dataset (2,200 samples, 22 classes), seven classifiers reach 97.1-99.6% under repeated stratified cross-validation, with Random Forest best. However, when Gaussian measurement noise of 0.25 feature standard deviations is added, Random Forest falls to 82.0% while SVM and KNN retain 92.9% and 92.3%, so the model ranking changes. On the Fertilizer Prediction dataset (100,000 samples, 7 classes) every model performs at chance (about 14.3%); statistical tests show the labels are independent of the features. We contribute a dataset-audit procedure, a noise-robustness protocol, and an open web prototype that warns when a model performs at chance.
Language: English
Subfield: Plant Science
Open Access status: green
OpenAlex ID: W7236300375
Cites:
Cited by:
Related to: 10