Visual Soil Analysis: an AI-powered solution

Abstract

This paper presents an image-based approach to soil classification using deep learning techniques. A custom dataset of 500 soil images, representing five distinct categories—Sandy, Loam, Alluvial, Clay, and Coarse—was compiled and used to train a YOLOv8 [1] object detection model. To enhance prediction reliability, a Python-based histogram correlation method was integrated, comparing RGB channel distributions between input and reference samples. The system outputs the predicted soil type along with a confidence score and similarity metric. Results demonstrate high classification accuracy and practical usability, delivered through a publicly accessible web application built with Streamlit [2], enabling real-time soil identification for researchers, students, and field professionals.