Solving the problem of recognition and segmentation of images of natural resources deposits based on an ensemble of neural networks

Abstract

Exploration of mineral deposits is a complex task, for the effective solution of which it is necessary to obtain the maximum possible amount of relevant information about the object under study. The article considers a possible solution to the problem of exploration of mineral deposits by using the tools of artificial neural networks. The main idea in building the architecture in this work was the ability to combine the segmentation of mineral deposits on the geological map of the region and their classification by type and volume. An ensemble of clustering algorithms has been developed to solve the segmentation problem. This ensemble provides a single vector of cluster labels describing the affiliation of geo-system points to certain clusters. To solve the classification problem, it was decided to use an ensemble of DeiT models that provide high-quality classification of hyperspectral images in the task of searching for ore deposits. The use of an ensemble of DeiT models has advantages over the use of individual DeiT models, in particular, this approach reduces the effect of retraining and improves the quality of classification.