A Stacked Deep Learning-Based Method for Multiclass Classification of Sandstones in Microtomographic Images

• Antônio Pedro V. Lima UFMA
• S. Italo Francyles UFMA
• Alan C. Araújo UFMA
• Aristófanes C. Silva UFMA
• Anselmo C. Paiva UFMA
• Deane M. Roehl PUC-Rio

Resumo

This study proposes a hybrid approach for the multiclass classification of micro-CT images of sandstones, aiming to support the automation of rock analysis in oil and gas applications, given that this process still heavily relies on manual analysis by specialists, which can be time-consuming, subjective, and error-prone. The methodology was developed using the 11 Sandstones: raw, filtered and segmented data dataset. Convolutional neural networks (DenseNet121, ResNet50, and VGG16) were trained individually, and their respective softmax outputs were weighted according to each model’s individual accuracy. The proposed method constitutes a stacked model by integrating the weighted averaging strategy with a multilayer perceptron (MLP). The results showed that the proposed approach outperformed both the individual models and simpler ensemble techniques, achieving up to 99% overall accuracy in the test evaluations. Furthermore, it proved effective in correcting errors observed in the other approaches, providing greater stability and robustness.