Design of SoilNet Framework to Classify Soil Types and Predict the Crop Yield Using Fusion Deep Learning Models

Abstract

Soil classification is a crucial aspect of precision agriculture, as it directly influences crop selection and yield optimization. Existing models lack the utilization of fusion models for classifying the type of soil, and hence, prediction is not successful. This study proposes SoilNet, a deep learning-based framework designed to classify soil images into four distinct categories. The model architecture includes five convolutional layers paired with max-pooling layers to extract hierarchical spatial features from soil images, followed by dense layers for classification. The Soil Image dataset, comprising RGB soil images resized to 220x220 pixels, was pre-processed through normalization and split into training and testing subsets. Hyperparameters, such as learning rate, batch size, and number of filters, were tuned to optimize the model’s performance. The experimental evaluation shows that the soil classification method achieved an accuracy, precision, recall, and F1-Score of 92.38%, 92.46%, 92.38%, and 92.18%, respectively. Comparative syntactic learners demonstrate that both feature extraction and parameter optimization using the method improve it, leading to a better fit of feature-intensified placing quality. A detailed analysis of the confusion matrix revealed only a few misclassifications, which verified the reliability and usefulness of the method. Introducing a new approach to automated soil classification, this study presents a scalable method to provide agriculturalists with efficient soil classification. Future work will focus on data augmentation, hybrid modeling strategies, and the real-time deployment of the approach for field use.