Segmentation of Microtomography Images of Carbonate Rocks: A Comparison Between Ordinary Kriging and Universal Kriging

• Maxwell Pires Silva UFMA
• S. Italo Francyles UFMA
• Aristófanes Corrêa Silva UFMA
• Anselmo Cardoso de Paiva UFMA
• Deane Roehl PUC-Rio

Resumo

This study proposes a hybrid approach for segmenting micro-CT images of carbonate rocks, combining the UNet convolutional neural network with geostatistical kriging methods (ordinary and universal). The process is divided into two main stages: image preprocessing (using CLAHE and histogram specification) and segmentation, where UNet detects the edges of the regions of interest, and kriging is responsible for filling these regions. Using a dataset of 16 carbonate rock samples from Brazil’s pre-salt region, the model was trained and evaluated using metrics such as IoU, Precision, Recall, and F1-Score. The combined use of CLAHE and histogram specification significantly improved UNet’s performance, achieving over 96% across all metrics. In the final segmentation stage, ordinary kriging with a 60% threshold outperformed universal kriging, yielding superior results in IoU (81%), Precision (95%), Recall (84%), and F1-Score (89%). These results indicate that integrating deep learning with statistical interpolation is effective, with ordinary kriging standing out due to its higher accuracy and lower computational cost. Future work includes exploring different trend functions and variogram models, as well as applying alternative neural network architectures and geospatial interpolation techniques.