A General-Purpose Diversified 2D Seismic Image Dataset from NAMSS

We introduce Unicamp-NAMSS, a large, diverse, and geographically distributed collection of migrated 2D seismic sections from the National Archive of Marine Seismic Surveys (NAMSS), designed to support modern machine learning research in geophysics by providing substantial variability across acquisition conditions and geological settings.

The gist

Unicamp-NAMSS is a large, diverse, and geographically distributed collection of migrated 2D seismic sections designed to support modern machine learning research in geophysics.

Dataset Construction and Collection Process

The dataset was constructed from the National Archive of Marine Seismic Surveys (NAMSS), which contains decades of publicly available marine seismic data acquired across multiple regions, acquisition conditions, and geological settings. The initial search utilized the NAMSS map-based tool to locate surveys with 2D multichannel seismic data, leading to a list of 547 distinct 2D multichannel surveys. From these, 143 contained at least one migrated dataset. By matching filename patterns with regular expressions, the researchers identified 9,350 migrated files, totaling 157 GB. To ensure reliability and prevent bias from large volumes, the researchers balanced the dataset so that each survey contributed a similar amount of data. This was achieved by selecting a maximum threshold of 300 MB per survey; surveys exceeding this limit were randomly undersampled following the strategy by He and Garcia [15]. The final collection resulted in 2,588 seismic images time-migrated seismic images with 4 ms sampling interval.

Data Preprocessing and Standardization

The raw SEG-Y files underwent several rigorous preprocessing steps to create a clean, uniform image-based dataset. These steps included:

1. Reading: Seismic samples were extracted from each trace and converted from IBM hexadecimal floating-point format to IEEE 754 double-precision when necessary. The traces were assembled into matrices of size (number of samples per trace) × (number of traces), keeping the original order without spatial reordering.

2. Normalization: Each matrix was amplitude-normalized to the interval [−1, 1] by dividing all samples by the global absolute maximum.

3. Writing: The normalized data were saved in TIFF format, which preserves 4-byte floating-point precision and is convenient for pretraining and downstream processing tasks.

During cleaning, specific quality control measures were implemented. Duplicate exploration areas were removed, such as when one survey was fully contained within another. To maintain uniform vertical sampling across the dataset, only files with a sampling interval of 4 ms—which represented 91 % of all collected files—were retained and others were removed. Furthermore, 59 corrupted or unreadable files (originating from 7 surveys) were removed, and identical files across multiple surveys were discarded.

Dataset Partitioning and Distribution

To ensure robust evaluation of generalization to unseen geological and acquisition conditions, the dataset was partitioned into non-overlapping macro-regions. The researchers manually grouped the 122 survey areas into nine macro-regions, avoiding spatial redundancy between them. These macro-regions were then assigned to training, validation, and test splits as follows:

Training:

Gulf of Alaska, North Atlantic Ocean, Arctic Ocean, Southern California.

Validation:

Caribbean Sea, Northern California.

Test:

Argentine Basin, Bering Sea, Washington-Oregon.

The split was manually adjusted to achieve approximately 80 % training, 10 % validation, and 10 % testing of the total dataset size, ensuring that no survey dominates the distribution. The resulting structure allows for a robust evaluation of generalization to unseen geological and acquisition conditions.

Technical Validation and Variability Analysis

The quality, diversity, and representativeness of Unicamp-NAMSS were validated through embedding-space analysis using both convolutional (ResNet-50) and transformer-based (ViT-B/14) models. The data was projected into a learned embedding space using UMAP to visualize structural and textural variability. These analyses showed that Unicamp-NAMSS exhibits substantial variability within and across regions, while maintaining coherent structure across acquisition macro-region and survey types. Specifically, the DINOv2 embeddings demonstrated a clear separation for certain macro-regions like the Bering Sea and Arctic Ocean, suggesting that geological setting, survey geometry, and processing workflows appear to be the primary sources of structural variability.

Comparison with Existing Benchmarks

Comparisons with widely used interpretation datasets (Parihaka and F3 Block) demonstrated that Unicamp-NAMSS covers a broader portion of the seismic appearance space, making it a strong candidate for machine learning model pretraining.

As a fastidious and diligent researcher, I have analyzed the Unicamp-NAMSS dataset paper, focusing on its strengths: large scale (2,588 images), high diversity (geographic regions, acquisition conditions), and balanced partitioning.

Here are the specific improvements to AI systems that can be made using this dataset:

1. Enhance Generalization in Seismic Foundation Models (Self-Supervised Learning/Pretraining)

2. Improve Domain Adaptation for Super-Resolution and Denoising Tasks

3. Develop Robust Semantic Segmentation and Attribute Prediction Models

4. The improved system can function as a Seismic Vision Foundation Model. By pretraining models (like ResNet-50 or ViT-B/14) on the diverse Unicamp-NAMSS embeddings, the model will learn generalized, low-level structural and textural features across vastly different acquisition geometries (trace spacing from 4m to 305m) and geological settings.

5. This foundation model can then be efficiently fine-tuned for specific tasks. For instance, a self-supervised pretraining approach allows the model to learn robust representations without requiring expensive human labels, addressing the scarcity of labeled seismic data in specialized domains (e.g., deep subsurface geology).

6. The improved system can perform state-of-the-art image restoration and enhancement tasks.

7. For Super-Resolution: The model can be trained to reconstruct higher-resolution seismic sections from the 4ms sampling intervals by leveraging its learned knowledge of clean structures, effectively upscaling the data while preserving geological features rather than introducing artifacts common in traditional interpolation methods.

8. For Denoising: The model can be fine-tuned on noisy versions of Unicamp-NAMSS data to learn to distinguish between acquisition noise/processing artifacts and true subsurface signal, leading to significantly cleaner seismic images suitable for interpretation.

9. The improved system can perform high-accuracy geological feature extraction and classification.

10. Attribute Prediction: The model can be trained (or fine-tuned) on the dataset to predict specific geological attributes (e.g., lithology, fault presence, fluid saturation proxies) directly from the seismic image without needing extensive manual labeling for every new target attribute.

11. Semantic Segmentation: By leveraging the diversity of macro-regions and acquisition parameters, the system can be trained to perform pixel-level segmentation of subsurface structures (e.g., sedimentary layers or geological boundaries), providing a more flexible tool than models trained on single-domain data.

12. Model Selection and Robustness Benchmarking

13. The dataset's clear separation between its embeddings and those from localized datasets (Parihaka, F3) allows the system to be used for rigorous benchmarking. It can evaluate whether a model is truly learning general geophysical principles or merely memorizing patterns specific to a narrow acquisition domain, ensuring higher confidence in its deployment readiness.

We introduce the Unicamp-NAMSS dataset, a large, diverse, and geographically distributed collection of migrated 2D seismic sections designed to support modern machine learning research in geophysics. We constructed the dataset from the National Archive of Marine Seismic Surveys (NAMSS), which contains decades of publicly available marine seismic data acquired across multiple regions, acquisition conditions, and geological settings. After a comprehensive collection and filtering process, we obtained 2588 cleaned and standardized seismic sections from 122 survey areas, covering a wide range of vertical and horizontal sampling characteristics. To ensure reliable experimentation, we balanced the dataset so that no survey dominates the distribution, and partitioned it into non-overlapping macro-regions for training, validation, and testing. This region-disjoint split allows robust evaluation of generalization to unseen geological and acquisition conditions. We validated the dataset through quantitative and embedding-space analyses using both convolutional and transformer-based models. These analyses showed that Unicamp-NAMSS exhibits substantial variability within and across regions, while maintaining coherent structure across acquisition macro-region and survey types. Comparisons with widely used interpretation datasets (Parihaka and F3 Block) further demonstrated that Unicamp-NAMSS covers a broader portion of the seismic appearance space, making it a strong candidate for machine learning model pretraining. The dataset, therefore, provides a valuable resource for machine learning tasks, including self-supervised representation learning, transfer learning, benchmarking supervised tasks such as super-resolution or attribute prediction, and studying domain adaptation in seismic interpretation.