An Efficient Deep Learning Framework for Soil Type Classification using Enhanced Image Preprocessing Techniques

Artificial intelligence techniques are increasingly utilized in agriculture, particularly for computer-assisted soil image classification. This work proposes an efficient and automated soil classification model that combines image preprocessing with deep learning to support applications such as crop selection, land management, and irrigation planning. The proposed model was trained and evaluated on a dataset of 1,563 soil images, categorized into four classes: clay, red, black, and alluvial. A robust preprocessing pipeline was designed to enhance critical texture features, including grayscale conversion, CLAHE-based contrast enhancement, HSV V-channel and RGB channel blending, followed by resizing and normalization. To improve generalization, transformations such as rotation and flip augmentations were applied. The proposed system utilizes a fine-tuned EfficientNetB0 as the feature extractor, integrated with a custom classification head consisting of global average pooling, batch normalization, dense layers with ReLU activations, dropout, and a softmax output. The model achieved a final testing accuracy of 97.94%, demonstrating the effectiveness of the proposed architecture for accurate and scalable soil type prediction in agricultural applications. Results demonstrate that the proposed approach outperforms traditional models, offering a scalable, low-cost, and accurate solution for precision agriculture.