Automatic Analysis Of Petrographic Thin Section Images Of Sandstone
A.Y. Bukharev, S.A. Budennyy, A.A. Pachezhertsev, B.V. Belozerov, E.A. Zhukovskaya, M.A. Tugarova, N.V. Bukhanov and A.N. Zakirov
Event name: ECMOR XVI - 16th European Conference on the Mathematics of Oil Recovery
Session: ECMOR XVI Poster Session 2
Publication date: 03 September 2018
Info: Extended abstract, PDF ( 6.12Mb )
Price: € 20
Petrographic data remains one of fundamental information sources for characterization of hydrocarbon reservoirs. Thin section analysis of sandstone aims to describe depositional textures, major grain types and granulometric distribution, sedimentary structures, grain sorting, mineral composition, structural features, pore types, porosity, etc. These features are exploited to interpret the sedimentary environments, to predict the distribution of sedimentary bodies and their geometry which determines a recovery factor and other key reservoir exploitation characteristics. To conduct these studies one is to segment thin section images first – to partition them into grains, fractures, cleavages, pores, cement. The segmentation process is time-consuming as it is carried out manually or with specialized software that requires a proper recipe preparation for each image. The segmentation accuracy directly shapes the quality of further petrographic analysis. The goal of work is to develop a fully automatic algorithm for segmentation of thin section images for sandstone and further analysis of partitioned objects. The developed algorithm combines both image processing (IP) and deep learning (DL) approaches. IP methods exploit color intensity and local textural information to segment key structural elements in thin section image: voids (pores and fractures). The combination of DL and IP methods exploit primary information from images to solve semantic and instance segmentation problems for grains and to classify grains, cement and pores. Implementing of DL approaches demands a comprehensive training sample, full enough to have a reasonable segmentation accuracy. Thereby, the dataset of labeled images has been prepared manually. The developed algorithm has been efficiently applied for thin section analysis of sandstone. It has showed not only high agreement with manually processed thin sections and tremendous working time optimization, but more consistent results of segmentation as well. The algorithm plays a role of auxiliary tool that simplifies significantly the petrographic analysis of sandstone: most routine processes are automated; each thin section specimen can be processed statistically in a straightforward manner.