Automation of Sediment Core Analysis: A Deep Learning Approach

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Nov 23 2023

In an article recently published in the journal Scientific Reports, researchers proposed a deep learning (DL)-based approach for automatic sediment core analysis.

Study: Sediment core analysis using artificial intelligence. Image Credit: dotshock / Shutterstock

Background

Understanding the subsurface stratigraphy is crucial for several societal and industrial applications, such as engineering geology and global climate change studies. Sediment cores act as the key source of information while investigating the subsurface, as it is inaccessible to direct observation.

Sedimentary facies, specifically packages of strata/sediment bodies formed in specific depositional environments, possess unique mechanical and physical properties that can be effectively utilized for subsurface stratigraphic modeling. Recent research demonstrated that constructing a detailed shallow subsurface model based on sedimentary facies properties can be an effective tool to predict damage risk from earthquakes and assess active tectonic deformation patterns. Thus, sediment facies analysis is the initial step in most environmental and Earth research studies. However, specific sedimentological training and expertise are necessary for high-resolution facies reconstructions. 

Artificial intelligence (AI) methods are increasingly being applied to environmental and Earth research in the past few years. However, the potential of DL and machine learning (ML) systems and the use of AI in geoscience have not been fully exploited in these proposed approaches. ML- and DL-based methods can play a critical role in automatizing the time-consuming sediment core analysis.

Visual performance of the model on five representative images of the validation dataset. The original full-resolution digital images, the model-produced segmentation masks, and the corresponding ground truths are shown in the left, central, and right columns, respectively.

The proposed approach

In this study, researchers proposed a novel method based on DL and convolutional neural networks (CNNs) to produce accurate sedimentary facies interpretations from standard digital images. Specifically, a DL-based approach was employed to perform automatic semantic segmentation on sediment cores digital images directly acquired in the field, leveraging the power of CNNs.

Semantic segmentation involves the classification of every image pixel based on a specific set of categories. A dataset of high-resolution digital images from continuous sediment cores of Holocene age that represent a broad range of continental to shallow-marine depositional environments was used in this study.

Additionally, researchers used six sedimentary facies associations as target classes in place of ineffective classification methods based only on lithology to maximize scientific value and optimize the interpretation process. Six Holocene sedimentary target classes were identified from Italy's Adriatic coastal plain and Po Plain of Apulia, Abruzzo, and Marche regions.

These classes include fluvial sand (FS), prodelta (P), peat layer (PL), swamp (Sw), poorly-drained floodplain (PDF), and well-drained floodplain (WDF) deposits, with an extra background class. Each core image was manually annotated by an expert sedimentologist, and a final dataset was created containing 82 high-resolution, non-overlapping digital images that were obtained from 32 continuous sediment cores with associated segmentation masks.

The dataset was divided into three mutually exclusive subsets, including a test set, validation set, and training set containing 12%, 11%, and 77% of the data, respectively. The performance of the model was evaluated on test and validation sets. Multiple standard segmentation metrics, including the balanced accuracy, F1-score, and the mean Intersection over Union (IoU), were measured during the model performance evaluation on the test and validation data.

Researchers generated the semantic segmentation mask of five full-resolution images from both test and validation sets. They compared them with the ground truths produced by the expert sedimentologist for visual model performance evaluation.

Significance of the study

A deep CNN was trained to automatically generate the semantic segmentation masks of digital images obtained from continuous sedimentary cores. The model represented a precise, easy-to-deploy, and fast tool that could significantly improve the subsurface stratigraphic modeling, making subsurface facies analysis accessible to professionals and scientists.

The mean IoU, F1-score, and balanced accuracy obtained by the model on the validation set were 0.884, 0.936, and 0.905, respectively, while the mean IoU, F1-score, and balanced accuracy on the test set were 0.853, 0.916, and 0.861, respectively. Thus, the scores obtained on both datasets did not display significant differences.

Additionally, the visual performance on validation data displayed a high correlation between the ground truths and model predictions. In a complex image in the validation dataset containing four target classes, including WDF, PDF, Sw, and PL, the model prediction accurately reproduced the sedimentologist segmentation mask through precise classification of most sedimentary facies.

However, minor errors localized at facies transitions, including Sw-PL, PDF-Sw, and WDF–PDF, were observed. No significant differences in visual performance were observed between the test and validation images. Four out of the five images used during the visual evaluation of model performance on test data were obtained from a set of sediment cores for which no images existed in the training set.

The objective was to evaluate the generalization capabilities of the model. In the image containing Sw, PDF, and WDF, all classes were accurately classified by the model on the test data, with satisfactory predictions for near facies transitions. Negligible errors were present, with a minor misclassification from PDF to Sw. In the other images, a near-perfect model prediction was achieved for all classes. In the image containing WDF, Sw, FS, and PDF classes, WDF was classified correctly, while minor errors were observed in Sw and PDF classifications. The P class was classified with high accuracy in both the validation and test datasets.