Deep Learning–Based Prediction and Classification of Minerals from Soil Images Using Computer Vision Techniques

Abstract

Precise determination of mineral composition of soils based on soil images is a very important issue in precision agriculture and in geospatial analysis as it directly affects crop productivity and land management procedures. The manual analysis however is time-consuming, error prone and cannot be scaled. To overcome these shortcomings, this paper presents a new deep learning-based system, DL-CVMIC (Deep Learning-Based Computer Vision for Mineral Identification and Classification), to efficiently and automatically classify minerals. The suggested model uses a multi-stage image processing with the steps of noise reduction, normalization, and contrast enhancement and then convolutional feature extraction to extract the spatial and textural features. They are then optimally classified using a hybrid deep neural network. The model is tested with high-resolution soil image datasets (512512 and 10001000) which show strong performance under different conditions. The experimental findings demonstrate that DL-CVMIC has a 96.8% accuracy in classification, a 95.9, 96.3 and 96.1 precision, recall, and F1-score, respectively. The classification is also much better with the proposed model compared to conventional methods like ANN (89.5%) and SVM (92.2%). Additionally, it reduces the misclassification rate to 3.2% and computational latency by 18.7%. These findings show that the suggested framework is scalable and stable to analyse soil minerals and provide decision support in real time in smart agriculture systems.