Abstract
Reliable property modeling is vital for Earth resource development, and seismic data can provide secondary variables to improve accuracy. However, seismic-integrated models remain uncertain due to inherent limitations in seismic data such as the cumulative effects of signal processing and attribute computation. In this study, we aimed to estimate a high-accuracy 3D secondary variable for porosity modeling from seismic attributes using a kriging-enhanced random forest (RF). This approach leverages the ensemble learning capabilities of RF to effectively handle limited training data, while incorporating the ability of kriging to account for spatial correlation. Prior to implementing this model, we developed an innovative workflow to correct seismic attributes based on geological trends. This workflow generated geologically informed seismic attributes by vertically correcting seismic attributes in areas of lower quality, while preserving their original lateral trends. We applied our methodology to a late Albian–early Turonian shallow-marine carbonate reservoir with a complex diagenetic history. After creating geologically informed seismic attributes, we used them, along with porosity well logs, as inputs for the kriging-enhanced RF model. This model calculated the mean of decision trees through kriging estimation rather than the usual averaging method. To evaluate effectiveness, we compared it with a deep neural network, a kriging-enhanced deep neural network, and a standard RF. The kriging-enhanced RF produced porosity closer to blind-well values than other methods and captured complex heterogeneities, such as channels and differing reservoir qualities across sequences, making the porosity cube a reliable 3D trend for further geostatistical simulations.