Fine-tuning pre-trained deep learning models for crop prediction using soil conditions in smart agriculture

Abstract

Agriculture is the backbone of the Indian economy, with soil quality playing a crucial role in crop productivity. Farmers often struggle to select the appropriate crop based on soil type, leading to significant losses in yield and productivity. To address this challenge, deep learning techniques provide an efficient solution for automated soil classification. In this study, a dataset of 781 original soil images, including clay soil, alluvial soil, red soil, and black soil, was collected from Kaggle and augmented to 3,702 images to enhance model training. Several deep learning models were employed for soil classification, including pretrained architectures and a proposed model, SoilNet. Experimental results demonstrated that DenseNet201 achieved 100% validation accuracy, ResNet50V2 98%, VGG16 99%, MobileNetV2 99%, and the proposed SoilNet model 97%. The proposed approach outperformed existing work by surpassing 95% accuracy. Additionally, model performance was evaluated using precision, recall, and F1-score, ensuring a comprehensive analysis of classification effectiveness. These findings highlight the potential of deep learning in improving soil classification accuracy, aiding farmers in making informed crop selection decisions.