Predicting Topsoil Suitability for Healthy Crop Production using a Threshold-Guided Deep Neural Network Model

Abstract

Healthy food production for the growing population usually affected by soil nutrient imbalance is a great challenge. This study presents a data-driven framework for predicting topsoil suitability for healthy crop production using physicochemical soil parameters and a Deep Neural Network (DNN) model. Soil samples were categorized using agronomic threshold ranges derived from international standards and national guidelines to group soils as suitable or unsuitable. A fully connected Deep Neural Network was implemented to discover the nonlinear relationships between soil attributes and suitability outcomes. To ensure robustness, the model was evaluated alongside some traditional classifiers using stratified cross-validation and independent testing. The Deep Neural Network achieved competitive performance on noise-free data and demonstrated superior robustness under increasing levels of Gaussian noise. The integration of threshold-based labelling with deep learning provides a practical pre-harvest decision-support tool. The proposed framework contributes to sustainable agriculture by enabling early detection of soil conditions that may compromise crop safety and productivity.