Enhanced ore classification through optimized CNN ensembles and feature fusion

Ore is a type of natural stone that contains economically valuable minerals or metals. Accurate classification of ore minerals is crucial for improving operational efficiency in mining, reducing environmental impacts, and determining market value. Traditional methods for classifying ores are often time-consuming, labor-intensive, and error-prone. Therefore, computer-aided systems offer a significant advantage in this field. In this study, various efficient Deep Learning (DL) approaches are utilized for the detection of ore types. Within the scope of the study, four different experiments (transfer learning, feature extraction and classification with SVM, feature selection with optimization algorithms, and ensemble methods) are conducted, and the methods are compared in terms of classification metrics. As a result of the experimental case studies, high accuracy rates between 95 and 98% are achieved. The most successful method is the ensemble method, weighted by grid search. The ensemble model, which combined AlexNet, VGG16, and Xception models, achieves remarkable results with an overall accuracy of 98.11%, precision of 98.18%, recall of 98.11%, and f1-score of 98.11% on the publicly available Ore Images Dataset (OID). This study demonstrates that efficient DL approaches can classify ores with very high accuracy and have significant potential applications in the mining industry.