• Confidence-weighted framework for on-site fertilizer quality screening • Uncertainty-aware off-spec detection using low-cost N–P–K sensors • Improved dilution robustness over rigid threshold-based decisions • Interpretable and lightweight for embedded deployment • Supports decentralized screening in smart agricultural systems Fertilizer quality plays a critical role in precision agriculture and smart farming systems, yet routine laboratory-based verification is often impractical for rapid decision-making at farm or plantation sites due to cost, time, and logistical constraints. Low-cost macronutrient sensors provide a promising alternative for decentralized fertilizer quality screening; however, sensor-based decisions are highly sensitive to measurement noise, dilution variability, and rigid threshold boundaries. This study proposes a confidence-weighted decision framework for on-site fertilizer quality screening and off-spec detection in smart agricultural systems using low-cost nitrogen (N), phosphorus (P), and potassium (K) sensors. The framework transforms macronutrient sensor readings into continuous nutrient-level confidence scores derived from laboratory-referenced threshold ranges and aggregates them through an interpretable weighting mechanism suitable for embedded implementation. Unlike conventional threshold-based classification, which enforces binary decisions, the proposed approach explicitly represents decision uncertainty and mitigates dilution-induced instability commonly encountered in field-deployable sensing scenarios. The framework was evaluated using fertilizer samples representing authentic, adulterated, and off-spec conditions across multiple dilution levels, with laboratory chemical analysis serving as ground-truth reference. Prior to decision framework evaluation, the low-cost macronutrient sensor was characterised metrologically: intra-class coefficient of variation (CV) values of 10.32%, 17.33–17.37%, and 18.24–18.49% were obtained for authentic, adulterated, and off-spec samples respectively (all within the 20% field-sensor acceptance threshold), and Pearson correlation with laboratory reference values yielded pooled r = 0.917 (R² = 0.840) for N, r = 0.814 (R² = 0.662) for P, and r = 0.848 (R² = 0.720) for K (all p < 0.0001). Experimental results on an independent 13-sample test set (39 measurements across three dilution levels) demonstrate that the proposed framework achieved an overall classification accuracy of 94.87%, outperforming the conventional threshold-based approach (92.31%), with a macro-averaged F1-score of 0.955 versus 0.933. The framework improved the classification accuracy of adulterated samples from 80.00% to 86.67% (recall: 0.867 vs. 0.800) and maintained stable classification outcomes for 92.3% of samples across dilution levels (Dilution Robustness Index, DRI = 0.962) with an average confidence margin of 1.59. Leave-One-Out Cross-Validation on the 36-sample training set confirmed a consistent generalization advantage (81.48% vs. 72.22%), while McNemar's test (χ² = 0.000, p = 1.000) indicates the accuracy difference is not statistically significant on this test set, motivating reliance on complementary robustness metrics. A lightweight Decision Tree baseline (depth = 3) achieved 100% test accuracy but only 83.33% LOO-CV accuracy and provides no uncertainty quantification, underscoring the advantage of the proposed approach for transparent, data-scarce field deployment. Notably, these robustness gains are achieved without reliance on data-driven machine learning models, preserving computational simplicity, transparency, and compatibility with low-cost embedded hardware platforms. The proposed decision framework supports pre-application fertilizer verification directly within smart agricultural workflows, enabling reliable, decentralised quality screening prior to field deployment and contributing to uncertainty-aware decision support in precision nutrient management.
Nugroho et al. (Wed,) studied this question.