Lithology Segmentation from Well Logs for Geothermal Exploration using Vision Foundation Models

Former oil and gas fields offer a repository of historical geophysical well logs that can help support geothermal exploration across large areas. Lithology classification from logging data is a fundamental task in subsurface geological interpretation. Existing deep learning approaches typically formulate this problem as a point-wise or sequence-wise classification task, where logging curves are treated as one-dimensional depth-dependent signals. Although such methods have demonstrated promising performance, they usually rely on large-scale labelled datasets for training. Moreover, logging datasets commonly exhibit severe class imbalance due to complex geological environments and strong heterogeneity, which further degrades the performance and robustness of data-hungry deep learning models.

To address these challenges, we propose a novel lithology segmentation framework, in which we reformulate lithology classification as a semantic segmentation task, where different lithological units are characterized by continuous intervals separated by distinct boundaries along the depth dimension. Based on this formulation, we develop a lithology segmentation framework that leverages large-scale vision foundation models, enabling effective learning under data-scarce and class-imbalanced conditions. Our core motivation is to transfer the strong image representation and generalization capabilities learned by large pretrained models on massive image data to the geological logging domain.

Specifically, well logging curves are transformed into two-dimensional pseudo-images by a structured multi-scale channel combination along the depth dimension. The repetition factor k controls how many times each logging curve is duplicated in the pseudo-image, enabling Vision Transformer (ViT) with fix-sized patches to encode logging patterns at multiple effective scales. For each scale k, a composite representation X^(K)∈ R^H×W_K  is formed by repeating selected logging curves with scale-dependent repetition factors, where H is the number of depth samples. Accordingly, the width of the pseudo-image at scale k is defined as W_k = k·N, where N is the number of logging curves. The final input representation X is obtained by concatenating all scale-specific representations: X = Concat(X⁽¹⁾, X⁽²⁾, X⁽²⁾, ..., X^(K)).

Building upon the pretrained Segment Anything Model (SAM), we retain the image encoder to extract high-level visual features, while a task-specific decoder is initialized and trained from scratch for lithology segmentation. The encoder weights are initially frozen and gradually unfrozen during training, and fine-tuned jointly with the decoder to adapt the feature space to the geological patterns of the specific domain. This staged training strategy stabilizes the optimization process, reduces overfitting with limited data, and effectively transfers knowledge from natural images to well logging images. Furthermore, by using a weighted loss function at the segmentation level to address class imbalance, it ensures that a minority of lithological classes contribute sufficiently to model updates.

Overall, the proposed framework demonstrates a new workflow for lithology interpretation by integrating foundation models with geological data analysis. It provides a data-efficient solution for lithology segmentation under realistic constraints of limited and imbalanced well logging datasets.