Interpreting Random Forest Crop Recommendations Using Soil and Climate Data

Authors

  • Purnama Wirawan Universitas Rokania
  • Ridwan Ridwan Universitas Rokania
  • Maruli Tua Universitas Rokania
  • Juliarni Juliarni Universitas Rokania
  • Mahmud Muzaki Universitas Rokania

DOI:

https://doi.org/10.47709/joa.v5i03.9880

Keywords:

Crop Recommendation, Data Audit, Feature Importance, Precision Agriculture, Random Forest

Abstract

High predictive accuracy in crop recommendation systems does not necessarily imply agronomic validity, particularly when soil–climate datasets contain class-linked patterns. Objective: This study assessed Random Forest crop-label classification while examining feature importance and within-class nutrient variability. We analysed a secondary dataset comprising 6,596 observations, 51 crop labels, and seven predictors (nitrogen, phosphorus, potassium, temperature, humidity, soil pH, and rainfall). We evaluated a 100-tree Random Forest using an 80:20 stratified train–test split, five-fold cross-validation, out-of-bag scoring, baseline comparisons, and eight random-seed trials. Interpretation included impurity-based and permutation importance, predictor-group ablation, correlations, and a class-level nutrient audit. Hold-out accuracy reached 99.70%, with a macro F1-score of 99.61%. Across eight seeds, mean accuracy was 99.8011% (sample standard deviation: 0.0987 percentage points). Nitrogen ranked first under both feature-importance methods, whereas soil pH ranked second under permutation importance. Nutrient-only and climate–pH models achieved accuracies of 79.70% and 95.23%, respectively. The audit identified constant within-class nitrogen, phosphorus, and potassium values in 29 classes covering 4,396 observations (66.65%), including two pairs of classes sharing identical nutrient triplets. High internal performance reflected discrimination within the supplied dataset, not validated suitability for new farming conditions. The findings highlight the need to verify data provenance, measurement units, and label construction and to undertake independent agronomic validation across farms and seasons before operational recommendations.

Downloads

Published

2026-11-09