Estimating soil fertility capacity through uncertainty management with neutrosophic logic

Abstract

This study proposes a machine learning framework that integrates Random Forest classification with neutrosophic logic for soil fertility prediction from 16 physicochemical properties. The model achieves a robust accuracy of 95% and incorporates a novel uncertainty quantification mechanism using neutrosophic values ​​of truth (T), indeterminacy (I), and falsity (F), calculated from prediction probabilities. The strategy assigns high confidence (T = 0.9) to probabilities ≥ 0.9, moderate indeterminacy (I = 0.3) to borderline cases (0.5 ≤ p < 0.9), and falsity (F = 0.5) to low-confidence predictions (< 0.5). This scheme enables tiered decision support, where 60% of cases receive automated recommendations, 25% require expert review, and 15% are flagged for rejection. Variable importance analysis reveals that clay (24.1%) and cation exchange capacity (CEC) (18.7%) are the dominant predictors, consistent with agronomic foundations, which strengthens the interpretability of the results. Overall, this approach contributes to the advancement of precision agriculture by complementing binary classifications with practical confidence metrics, particularly useful for managing uncertain soil condition scenarios.