Machine Vision for Soil Health Monitoring: Texture-Based Fertility Prediction Using Deep Learning

Abstract

Effectively evaluating soil fertility is key to maximizing rice yield. Since soil texture greatly influences water retention, nutrient availability, and root growth, our study introduced a deep learning framework for analyzing soil texture images. This framework aims to equip farmers with a reliable tool for better agricultural decision-making. We compared various convolutional neural network architectures, including traditional models like LeNet-5 and AlexNet, as well as more advanced hybrids, along with our custom design, the SoilTex-MSCNN, which incorporates a Multi-Scale Feature Block. This custom model showed exceptional performance, achieving an accuracy of 94.70% by effectively recognizing multi-scale texture patterns. Additionally, a hybrid MobileNet combined with LSTM also performed strongly, significantly outperforming earlier models. The findings clearly demonstrate that advanced, specialized CNN architectures are highly capable of distinguishing subtle soil textures, enabling accurate fertility assessments that support site-specific nutrient management and resource efficiency. Ultimately, this advances more sustainable and productive rice farming.