Soil Classification Using Hybrid Models: A Comparative Analysis

The soil-based tasks such as farming and construction, are rapidly increasing with the growth of industries and urbanization. Accurate soil classification helps farmers choose the right crops for better yields and ensure safe and stable building foundation in construction. Using a machine learning model can simplify the work and save time. However, these models often face challenges when dealing with image data, particularly with feature extraction which can impact their overall performance. This study focuses on overcoming this issue by using a hybrid approach combining deep learning models, such as Convolutional Neural Network (CNN), with traditional models like Support Vector Machine (SVM), KNearest Neighbors (KNN), and Random Forest (RF). The Kaggle soil image dataset, containing 1215 images is resized and normalized to get a standardized dataset. Features are extracted from these images using the ResNet-50 model, a deep learning architecture known for capturing important features in images. The models mentioned above are used for the classification. The hybrid models like SVM+CNN, KNN+CNN, and RF+CNN are trained using the features and their performance is validated with stratified k-fold cross-validation. The SVM+CNN hybrid model outperformed the rest with an accuracy of 97%. This high accuracy shows how combining deep learning and traditional machine learning can improve soil classification.