AdaSemSeg: An Adaptive Few-shot Semantic Segmentation of Seismic Facies

TL;DR

AdaSemSeg tackles the problem of semantic segmentation of seismic facies under limited supervision and varying class counts across datasets. It decomposes multi-class segmentation into multiple binary tasks processed by a single shared DGPNet backbone, leveraging Gaussian-process regression in latent space to condition predictions on class-specific support masks. The image encoder is initialized with self-supervised representations learned from seismic data (SimCLR), enabling effective feature extraction without large annotated datasets. Evaluated in a leave-one-dout setting on three public seismic facies datasets, AdaSemSeg shows competitive performance to target-trained baselines and outperforms prototype-based FSSS and transfer-learning approaches, demonstrating strong cross-dataset generalization and practical utility for adaptive seismic facies interpretation.

Abstract

Automated interpretation of seismic images using deep learning methods is challenging because of the limited availability of training data. Few-shot learning is a suitable learning paradigm in such scenarios due to its ability to adapt to a new task with limited supervision (small training budget). Existing few-shot semantic segmentation (FSSS) methods fix the number of target classes. Therefore, they do not support joint training on multiple datasets varying in the number of classes. In the context of the interpretation of seismic facies, fixing the number of target classes inhibits the generalization capability of a model trained on one facies dataset to another, which is likely to have a different number of facies. To address this shortcoming, we propose a few-shot semantic segmentation method for interpreting seismic facies that can adapt to the varying number of facies across the dataset, dubbed the AdaSemSeg. In general, the backbone network of FSSS methods is initialized with the statistics learned from the ImageNet dataset for better performance. The lack of such a huge annotated dataset for seismic images motivates using a self-supervised algorithm on seismic datasets to initialize the backbone network. We have trained the AdaSemSeg on three public seismic facies datasets with different numbers of facies and evaluated the proposed method on multiple metrics. The performance of the AdaSemSeg on unseen datasets (not used in training) is better than the prototype-based few-shot method and baselines.

Figures (19)

• Figure 1: In this figure, we illustrate the task of predicting the seismic facies (a multi-class segmentation mask) in a query image from an unseen seismic dataset f3_facies_netherlands_2019 by a semantic segmentation model using a few annotated examples from the dataset. This presents a realistic scenario for the interpretation of seismic facies. We address this problem using the few-shot semantic segmentation method.
• Figure 2: This figure illustrates the heterogeneity in the number and types of facies across seismic datasets. In this example, we show the classification of the horizons of the Parihaka Parihaka_facies_new_zealand_2020 and Penobscot Penobscot_facies_canada_2021 datasets, where we observe a difference in the number of seismic facies and variation in the naming convention of the facies.
• Figure 3: Tasks were generated using samples in the seismic datasets for the meta-training and meta-testing stages in the FSSS methods. $I^S$ and $M^S$ represent the support images and corresponding masks used in both the meta-training and meta-testing stages. Similarly, images in the query set ($Q$) are represented as $I^q$ and the associated masks as $M^q$. In this illustration, we use $K=5$ support images and $1$ query images from the F3 f3_facies_netherlands_2019 and Parihaka Parihaka_facies_new_zealand_2020 facies datasets in the meta-training stage. A similar setting is used for the Penobscot Penobscot_facies_canada_2021 dataset in the meta-testing stage. Tasks in the meta-training stage are used to train the parameters of the FSSS model, and the trained model is evaluated on tasks in the meta-testing stage.
• Figure 4: A schematic of the few-shot semantic segmentation (FSSS) algorithm. We show here how a task in the meta-training stage is used to train the model parameters. The FSSS method uses $5$ support examples (i.e., $K=5$ shots) from the F3 facies dataset f3_facies_netherlands_2019 to make multi-class prediction (six classes) on the query image. The forward passes in the model are represented using blue arrows. The predicted mask is compared with the ground truth to compute the loss that is used to update the model parameters. The backward passes allowing the flow of gradients are represented with red dashed arrows.
• Figure 5: The architecture of the DGPNet dgpnet_2022 used for the binary segmentation task. The DGPNet uses $K=5$ support examples to predict the binary segmentation mask for a single facies type in the F3 facies dataset f3_facies_netherlands_2019 on the query image. The forward passes in the model are represented using blue arrows. The predicted mask is compared with the ground truth to compute the loss that is used to update the model parameters. The backward passes allowing the flow of gradients are represented with red dashed arrows.