Machine Learning-Driven Soil Health Analysis for Precision Agriculture: Sensor Based Fertilizer Recommendation

Abstract

Optimal application of fertilizer is key in maximizing crop yields with minimal damage to the soil health, yet conventional soil health analysis remain labour intensive, time consuming and inaccessible to many farmers. This study proposes a machine learning-based fertilizer recommendation system utilizing soil nutrient status to improve soil health analysis. A hybrid machine-learning model synthesizing Extreme Gradient Boosting and Random Forest was trained on a dataset containing Agronomical fertilizer recommendations based on Nitrogen, Phosphorous, Potassium, Soil type and crop type. The study evaluated the performance of an RS485-based digital NPK soil sensor in a sub-study by correlating sensor readings with laboratory test results. In 10 samples per soil type, sensor readings were found to be within ±10% of laboratory values, proving their reliability in real-time field measurements. The findings identify the potential for combining ensemble machine learning algorithms with low-cost sensors to offer scalable, real-time, and accurate fertilizer recommendations for smallholder and commercial farmers. The integrated approach supports data-driven agricultural decision-making and opens the way for smart, sustainable nutrient management.