DATA-DRIVEN SOIL HEALTH ASSESSMENT USING GRADIENT BOOSTING AND DEEP NEURAL ARCHITECTURES

Abstract

Soil health is a fundamental factor for sustainable crop production and affects crop yield and ecosystem equilibrium. Conventional soil inspection techniques are tedious and time-consuming, prompting the development of efficient data-driven solutions. To facilitate soil quality classification in our research, we propose a deep learning-based soil quality classification method using extreme gradient boosting (XGBoost) and deep neural network models, which are multilayer perceptrons (MLP) and artificial neural networks (ANN). A dataset of soil samples characterized for pH, electrical conductivity (EC), phosphorus, potassium, organic carbon (C), and lime content was extended by controlled perturbations to increase the generalization of the model. The performance of the models was assessed based on accuracy, precision, recall, F1-score, and using ROC curves, and calibration plots. XGBoost provided a maximum accuracy of 94.25% and outperformed deep learning models (ANN: 74.75%, deep MLP: 68.25%, simple MLP: 60.25%). This suggests that machine learning can be successfully employed to automate soil health evaluation and aid data-based decision-making in precision agriculture.