The identification and quantification of organic macerals of hydrocarbon source rocks based on the improved U-net model

The identification and quantification of organic macerals of hydrocarbon source rocks are the main contents of organic petrological research. However, the organic matter content in hydrocarbon source rocks is low and highly dispersed. The traditional method of point-counting statistics by microscopic observation has characteristics of high work intensity and low efficiency, which restrict the further promotion and application of the traditional method. Based on the characteristics of microscopic images, this study proposes an intelligent quantification technology by using the optimized U-net model in which the dual attention transformer (DAT) and the decoder-guided recalibration attention (DRA) are introduced. Compared to the traditional U-net model, this method can be used to relative accurately capture the detailed information in the images, with relatively precise segmentation effect. On this basis, 1000 sets of fluorescence and white light reflection photos under the same field of view were trained with fourtypes of labels: sapropelite, exinite, vitrinite, and inertinite, to segment the pixel distribution ranges of the corresponding components from those photos. The OpenCV library was used to count the number of pixels of various components and to calculate the corresponding proportions, and then to calculate the TI index in order to determine the type of organic matter, thus to achieve the artificial intelligence-based quantitative statistics of organic macerals. At the same time, by introducing the transfer learning technology, the accuracy and robustness of the model were improved, with the achieved model′s mean Dice (mDice) of 93.6% and the achieved mean intersection over union (mIOU) of 90.2%. The samples with different organic matter contents were selected as the validation set. By comparing the systematic test results with the manual recognition results, it was found that the content level and dispersion degree of organic matters did not affect the accuracy of segmentation and recognition. The relative errors were controlled between –5% and 5%, and the differences in contents of each kind of components were between –2% and 2%. The experimental results are highly consistent with the statistical results of the artificial point-counting method, which verified the feasibility of the model. This research is expected to promote the development and application of computer vision technology in the field of organic petrology.