Enterprise AI Analysis

Key Findings: XAI Visualizations and Model Interpretation

The study successfully applied XAI (Grad-CAM heat maps) to a deep-learning model for bioturbation intensity classification. The visualizations demonstrated strong alignment with experienced ichnologist interpretations, confirming the model's ability to focus on relevant geological features while disregarding artificial markings. For unbioturbated strata, heat maps highlighted continuous sedimentary structures. For moderately bioturbated strata, the model accurately identified individual trace fossils or disruptions of sedimentary structures. In intensely bioturbated strata, the heat maps reflected pervasive bioturbation or localized areas contributing most strongly to the classification.

Implications and Future Work for XAI in Geoscience

The integration of explainable AI (XAI) into geological studies represents a significant advancement in bridging the gap between artificial intelligence and traditional geoscientific interpretation. One of the key challenges in applying AI models to geological problems is their "black-box" character, referring to the lack of transparency in how these models process inputs and generate outputs. This commonly leads to some level of skepticism among geoscientists who rely on intuitive, experience-based decision making. By incorporating XAI techniques, geologists can gain deeper insights into how AI models reach their conclusions. This transparency enhances trust in AI-driven predictions, as geoscientists can validate whether the AI is recognizing meaningful geological features, such as burrows, grain sorting, or sedimentary fabric, rather than irrelevant noise within the data.

From an ichnological perspective, the three-class bioturbation classification adopted in this study is useful for capturing general variations in bioturbation intensity and for distinguishing broadly between low, moderate, and high degrees of sediment reworking. This simplified scheme is particularly effective for automated image-based analyses, where subtle distinctions between adjacent bioturbation classes may be difficult to resolve consistently and can be subject to significant interpretive uncertainty even in manual assessments. Nevertheless, we acknowledge that the development of a six-class bioturbation intensity model, explicitly aligned with the full Bioturbation Index (BI) framework 20, could provide additional interpretive value. A full bioturbation intensity classification would allow more detailed characterization of ichnological variability, enhance sensitivity to subtle environmental changes (e.g., offshore versus prodelta), and improve applicability in complex depositional settings where the full spectrum of bioturbation intensities is expressed e.g., 25,26,27,28. Such an approach would also facilitate more direct comparisons between AI-derived outputs and traditional ichnological analyses.

It is also important to mention that a wide range of siliciclastic depositional environments have been covered in the training data set 7. However, misclassification may still arise from limitations in image quality and geological context. For example, very fine-scale (diminutive) or cryptic bioturbation that fall below the effective pixel resolution of the image, or individual trace fossils that are larger than the image dimension may be underrepresented or misclassified. Images of bioturbation of carbonate-rich systems may look different in terms of trace fossil diversity or sedimentary structures than the training data set used in this study, which may result in misclassifications. Low image resolution, uneven lighting, motion blur, or out-of-focus images can obscure biogenic fabrics and reduce model confidence. Irregular outcrop surfaces, unnatural marks, and shadows from uneven surfaces can introduce geometric distortions that are poorly represented in the training data, potentially leading to erroneous predictions. Therefore, although this automated approach may reduce interpretation time, results should be validated by an expert before making a final decision. For optimal application of the algorithm, using high-quality images acquired perpendicular to surfaces, with consistent lighting and minimal shadows are recommended. Future studies incorporating multi-expert annotations and quantitative agreement metrics will be essential for fully benchmarking AI-derived bioturbation interpretations against expert judgment and for improving reproducibility across depositional settings.

Methodology: Dataset and Grad-CAM Application

The dataset used in this study was previously introduced by Ayranci et al., 7. It includes 262 test images with 233 correctly classified bioturbation intensity classes (unbioturbated, moderate, and intense), as well as 29 misclassified images. The bioturbation intensity labels were assigned, following established ichnological frameworks (e.g., the bioturbation index 20) by an experienced ichnologist. The images with no bioturbation include zero or negligible bioturbation (Fig. 12). These images dominantly display well-preserved primary sedimentary structures including planar parallel lamination, cross lamination and bedding, fractures, rip-up clasts, concretions, broken core pieces, pen-marks on cores, scattered pebbles, mudstone drapes, variations in lithology (e.g., sandy and muddy intervals), saw marks, soft-sediment deformation structures, hummocky cross-stratification, current ripple cross-lamination, gravel lags, and erosional truncation surfaces.

Images of moderately bioturbated strata display physical sedimentary structures (similar to those observed in images of unbioturbated strata) coupled with moderate degrees of bioturbation represented by individual trace fossils and disruptions in primary sedimentary structures (Fig. 12). These images either display homogeneously moderate bioturbation throughout the images or localized intervals of bioturbation.

Images of intensely bioturbated strata are characterized by homogeneously intense bioturbation with very little or no preserved sedimentary structures (Fig. 12).

There are a few tools to apply XAI in image classification, including Grad-Cam, LIME, and Integrated Gradients. Given that the Grad-Cam method has been reported to be one of the fastest and easiest methods 29,30, it was also the selected method for this study. This method provides class-specific, spatially faithful visualizations that align with Convolutional Neural Network's (CNN) internal mechanisms, ensuring faithful, interpretable, and computationally efficient explanations. By contrast, general-purpose methods like LIME and SHAP are less reliable for structured image data, owing to perturbation instability, independence assumptions, and computational inefficiency 31.

The technique begins with identifying the final convolutional layer in the model, because it retains the spatial information that is crucial for localization. Then, we performed a forward pass through the network with the input image to obtain the logits (class scores) for each class in the final output layer. At this point, the score corresponding to the class for which we want to generate the activation map should be selected. The gradient of the selected class score with respect to the feature maps of the final convolutional layer is computed, highlighting how changes in these activations affect the class score. We then apply global average pooling to these gradients to derive weights indicating the importance of each feature map. These weights are used to combine the feature maps into a class activation map, wherein each feature map is weighted by its corresponding importance. Finally, the class activation map is up-sampled to match the dimensions of the input image, ensuring that the localization results can be overlain on the original image for visualization (Fig. 13).