Interpretation of Computer Vision Models in Mapping of Opencast Mining Sites: Overview
A. A. Kolesnikov, A. V. Reznik, A. A. Belosludtseva, V. L. Gavrilov, K. E. Medvedeva
Chinakal Institute of Mining, Siberian Branch, Russian Academy of Sciences, Novosibirsk, Russia
Keywords: Model interpretation, satellite images, machine learning, image segmentation, opencast mining, production, disturbed land
Abstract
This article overviews the methods of interpretation of machine learning used in image segmentation of opencast mining areas. The model checkup used the information on mining-disturbed land in the Novosibirsk Region and in some districts of Yakutia. Examples of the formation and time history of such land outlines are given. The scope of the analysis embraces various approaches to interpretation of computer vision modeling, including Grad-VAM, Saliency Maps, Feature Maps, Occlusion Sensitivity, Integrated Gradients and Counterfactual Explanations. The studies demonstrate advantages and disadvantages of each method, and imply that an increase in the number of learning iterations not always brings improvement in the model quality. The authors advise on selecting initial data interpretation techniques with regard to peculiarities of an observation target when solving certain geotechnical or geoecological problems.
|
