Estimation of physico-chemical properties of soil using machine learning

Abstract

Alternatives to traditional soil quality estimation methods, often costly and inaccessible in resource-poor regions, are needed. This study explored the use of soil images to determine physico-chemical characteristics of soil samples from two locations in Benue State, North-Central Nigeria: Vandeikya (7.099653°, 8.572102°) and Konshisha (6.783539°, 9.069549°). The study integrated Machine Learning (ML) and Deep Learning (DL) techniques with soil image analysis, employing image feature extraction through Gray-Level Co-occurrence Matrix (GLCM) and Gabor filters to enhance the feature space across 1,388 samples (each consisting of 600 g of soil) and 16 properties. We evaluated different models, including Support Vector Regression (SVR), a Convolutional Neural Network (CNN), an optimized CNN using Grid Search, and a stacked ML-CNN approach. The models were assessed on their ability to predict soil properties while minimizing error rates. The SVR model showed limited predictive performance, with its best results yielding slightly over 40 % R² and negative scores for physical properties. The CNN model performed better but had feature scores below 80 % R². The optimized CNN model surpassed 90 % R² and reduced MAE by more than 50 % for several properties. The stacked ML-CNN model further improved these metrics, demonstrating the effectiveness of the integrated approach. This research highlights the potential for using machine learning and image-based analysis as a scalable, cost-effective alternative to conventional soil testing, offering practical solutions for regions where conventional measurement tools are scarce. Integration with historical soil data might also open new possibilities for enhancing the accuracy and applicability of these methods for broader soil management practices.

Keywords

• Soil properties
• Soil images
• Machine learning
• Image analysis
• Agriculture