SoilNet – Soil Texture Classification using Convolutional Neural Network

Abstract

Soil texture classification is a critical task in agriculture, as it directly impacts water retention, nutrient availability, and crop planning. Traditional approaches rely on manual inspection and laboratory analysis, which are time-consuming and subject to human error. This study introduces SoilNet, a Convolutional Neural Network (CNN)-based model designed to automate soil texture classification using image data. The proposed architecture consists of four convolutional blocks for hierarchical feature extraction, followed by a fully connected layer that outputs class probabilities for three texture categories: coarse, medium, and fine. The model was trained and evaluated on a combined dataset of 3,702 soil images from Roboflow’s SOIL (v1) dataset and additional Kaggle samples. After hyperparameter tuning - optimizing epochs, learning rate, kernel size, and activation function - SoilNet achieved an outstanding 99.46% accuracy, 0.0182 test loss, and near-perfect precision, recall, and F1-scores. These results demonstrate the model’s robustness and its potential as a reliable tool for rapid, scalable, and objective soil texture classification in precision agriculture.