A Machine Learning Framework for Precision Soil Quality Characterization for Crop-Specific Fertilizer Recommendation in Heterogeneous Agro-Ecosystems

Abstract

Accurate fertilizer recommendation is fundamental to precision agriculture, yet many machine learning models rely on static soil nutrient measurements without adequately incorporating environmental factors that influence nutrient availability. This study proposes a hybrid machine learning framework that integrates soil chemical, physical, and climatic variables into a Principal Component Analysis (PCA)-derived Soil Quality Index (SQI) for crop-specific fertilizer recommendation. The framework combines PCA-based feature extraction with Random Forest, Gradient Boosting, and Multilayer Perceptron (MLP) classifiers. Experiments were conducted on an 8,000-record dataset comprising 11 crop types, five soil classes, and seven fertilizer categories. Model development included standardized preprocessing, stratified train-test partitioning, and exhaustive hyperparameter optimization. Despite this rigorous framework, all models achieved only 12.4–14.0% test accuracy, equivalent to the 14.29% random baseline for a seven-class classification problem. Statistical auditing revealed that fertilizer labels were not significantly associated with soil or crop types (Chi-square, p = 0.584 and p = 0.100, respectively), while mutual information between continuous predictors and fertilizer labels was negligible. Learning-curve analysis further showed that the models have excellent training performance but no improvement in validation accuracy, indicating poor predictor-label dependency rather than inadequate model capacity. These findings demonstrate that dataset quality, rather than algorithm selection, constrained predictive performance. The primary contribution of this work is a robust and reusable machine learning pipeline complemented by a systematic predictor-label dependency auditing protocol. This framework provides a practical methodology for validating agricultural datasets before model deployment, thereby improving the reliability, reproducibility, and trustworthiness of precision agriculture and digital soil mapping applications.