Comparison of supervised and self supervised approaches for micro-CT lithology classification of carbonate rock samples

Resumo

The characterization of pre-salt reservoirs is a challenging task in the oil industry due to the geological
peculiarities and the heterogeneity of carbonate rocks. These challenges gave rise to new methods in order to better
characterize these rocks such as computed tomography for inner structure image generation and new computational
methods to analyse them. One of such methods is the use of artificial intelligence techniques, such as deep learning
that is considered the state of the art in several tasks and specially on computer vision. This work employs deep
learning techniques for the lithological classification of rock samples in micro-tomography images of cylindrical
rock samples referred usually as plugs. Two training paradigms are tested and compared, namely, supervised
and self-supervised training. The experiments employed densenet161 pretrained on ImageNet as base model,
which is a notorious model for image classification on the Imagenet dataset. The contrastive learning method
called supervised contrastive (supervised adaptation of SimCLR) was chosen for the self-supervised experiments.
This method allows the use of the label information in the SimCLR loss function while also enabling the authors
to incorporate this information in three different ways: the standard SimCLR framework (where each image is
considered unique), label-based (where each sample belonging to the same lithological classification is considered
equal) and sample-based (where all images generated from the same rock sample are equal). The dataset consists of
46,185 images from 623 rock samples which are distributed between seven different classes labeled by specialists.
Experiments with different training sizes were performed and compared for the supervised and self-supervised
cases. The results obtained over 70% accuracy and F1 score on all experiments with standard deviations equal
or lower than 7 % in most cases. The supervised experiments achieved the best results but the self-supervised
approaches also displayed comparable results.