<p>A key aspect of large scale mining is an accurate identification of rocks and ores. This study focuses on classifying various rock types, including igneous (basalt, granite), sedimentary (coal, limestone, sandstone, shale), and metamorphic (marble, quartzite), along with ores such as iron (hematite, magnetite, limonite, siderite), copper (chalcopyrite), aluminum (bauxite), and lead (galena). Data are collected from various locations in India. The classification process involves two steps: grain boundary detection using cellular automata (CAs) and classification using convolutional neural networks (CNN). A proposed cellular automaton (CA) based algorithm enhances grain visibility and generates refined images that serve as input to the CNN models. Two CNN models, referred to as Model-I and Model-II, are proposed for rock and ore classification, respectively. To enhance feature extraction and improve classification accuracy, a self-attention module is incorporated before the fully connected layer in Model-II. Comparative analysis shows Model-I achieves high accuracy for rock classification (97.8%) but performs moderately for ores (91.1%), while Model-II performs well for both rocks (97.9%) and ores (94.12%), however, requires significantly more computational time. Comparative analysis suggests Model-I is more efficient for rock classification, whereas Model-II is better suited for ores. Validation with an independent dataset confirms the reliability of both models. Moreover, the validation of the models is confirmed by comparing our models with five existing deep learning models, 2 pre-existing related works and three attention mechanisms with Model-II. To improve accessibility, the models are deployed on Amazon Web Services (AWS) cloud, and an Android application is developed. Users can upload rock or ore images via the app, which processes them through the cloud-hosted model for real-time classification, making the system practical for field applications.</p>