Estimation of the dolomite content of carbonate rock outcrops based on spectral knowledge and machine learning

Abstract

Accurately estimating the dolomite content in carbonate rocks is crucial for optimizing oil and gas exploration and production strategies. Hyperspectral techniques for estimating dolomite content have advantages in terms of efficiency, cost-effectiveness, and non-destructiveness compared with traditional laboratory methods. Despite the abundance of hyperspectral data, feature selection and extraction remain challenging. In this study, hyperspectral data collected from surface outcrop in the field using the analytical spectral device (ASD) were applied to construct model for estimating dolomite content. Firstly, the data were preprocessed via outlier analysis and continuum transformation. Next, a hybrid approach integrating spectral knowledge with machine learning was proposed and applied to facilitate efficient and precise feature selection of the hyperspectral data; in this approach, preliminary screening based on spectral knowledge is followed by further hyperspectral data feature selection using a random forest algorithm. The selected features were then combined using a support vector regression algorithm to obtain the estimation model. Finally, the accuracy of the model was evaluated using the hyperspectral data from field outcrop samples. To further verify the effectiveness of this method, various combinations of eight input variables and four machine learning algorithms were compared. Among all combinations, our model achieved the highest accuracy with a test R2 value of 0.91 and a root-mean-square error of only 0.122. The proposed method is practical and efficient and provides precise quantitative data for field geologists to identify the mineral distribution in outcrops. Thus, our method provides robust support for understanding reservoir characteristics and has significant practical value in geological surveys and mineral exploration.