Interpretação de zonas carstificadas usando redes neurais convolucionais e interpretabilidade através de Explainable AI

Abstract

Ground penetrating radar (GPR) is a geophysical tool that can be used to assist in mapping karstified zones in analogs for the characterization and understanding of carbonate reservoirs. With the aid of GPR, it is possible to understand the behavior of the karstification processes in carbonates and, thus, expand the knowledge to the reservoir level, as well as make parallels with analogues in the Brazilian pre-salt. In addition, Machine Learning (ML) and Deep Learning (DL) algorithms have allowed the application of computer vision techniques to identify geological structures and facies based on geophysical data, as obtained through the GPR tool, but mainly with based on seismic. In this context, and using as data eight GPR sections, and five attributes generated from them (energy, similarity, instantaneous phase, instantaneous frequency, and the ratio between the Hilbert trace and the similarity), this study seeks to apply DL models based on convolutional neural networks using the U-Net architecture. Moreover, Explainable Artificial Intelligence (XAI) techniques using SHapley additive exPlanation (SHAP) values are applied to improve the interpretability and explainability of the generated models. These techniques were employed in order to assess the rules found by the models, the modeling quality and the presence of biases in the model. Moreover, distinct settings with regard to background SHAP values were tested and compared to assess how they influence model explainability. As demonstrated in the results, the U-Net architecture developed was able to perform image segmentation through GPR data and, consequently, map and differentiate karstified from non-karstified zones. Furthermore, the SHAP values show that the energy attribute was the feature that provided more information in the modeling and consequently provided a greater weight in the model rules while the other features presented less relevant contributions. Furthermore, the type of sampling used to define the reference values for the SHAP values resulted in different interpretations of the contributions of the features. Finally, the present work generated models capable of assisting in the mapping of karstified zones, further supported by a technique capable of promoting the understanding of complex models and allowing greater cooperation between experts in geosciences and the results generated through ML and DL techniques.