The geosciences are facing a data deluge. Modern seismic surveys and continuous global monitoring networks generate terabytes to petabytes of data, capturing everything from high-resolution 3D reflections of the Earth's crust to continuous microseismic chatter. Historically, interpreting this data has been a manual, artisan craft. Geophysicists would visually inspect seismic profiles to trace geological faults, outline salt domes, pick wave arrival times, and filter out noise.
As noted across the field—from structural mapping to earthquake monitoring—this manual approach has hit a hard scalability bottleneck. Human expert review is slow, subjective, and entirely unable to keep pace with continuous data streams. To overcome this, a major research direction has emerged: reframing seismic interpretation tasks into computational frameworks, leveraging computer vision, signal processing, and deep learning to automate subsurface feature mapping and event detection.
Reframing Subsurface Volumes as Natural Scenes
One of the core intuitive leaps in automating seismic interpretation is treating 3D subsurface data volumes not merely as collections of independent acoustic traces, but as highly structured visual scenes or complex textures.
When a human interpreter looks at a seismic slice, they naturally group regions with similar wave patterns and focus on sharp discontinuities. To replicate this computationally, researchers first turned to texture attributes and visual signal processing. Zhiling Long evaluated how computer vision texture descriptors, such as Local Binary Patterns (LBP) and Gray-Level Co-occurrence Matrices (GLCM), could automatically segment complex geological structures like salt domes and faults. By treating the subsurface as a texture classification problem, these algorithms can map large-scale structures far faster than manual tracking.
Expanding on this perspective, Mohamed A. Deriche et al. formalized a roadmap for treating seismic volumes as "natural scenes." In Subsurface Structure Analysis Using Computational Interpretation and Learning: A Visual Signal Processing Perspective, the authors demonstrated that computer vision techniques like 3D texture gradients and tensor-based tracking can reliably isolate critical geological boundaries. Because geological structures manifest at multiple scales, Motaz Alfarraj et al. introduced multiresolution analysis (MRA) using directional curvelet transforms in Multiresolution analysis and learning for computational seismic interpretation. This allows algorithms to analyze seismic textures across various orientations and scales simultaneously, mirroring the macro-to-micro evaluation performed by human geologists.
[Raw 3D Seismic Volume]
│
▼
[Directional Curvelet / Texture Transforms] ──► Captures multi-scale orientations
│
▼
[Saliency / Attention Mapping] ───────────────► Mimics human visual focus
│
▼
[Automated Facies / Fault Segmentation]
Other approaches directly mimic the human visual system's focus mechanisms. Muhammad Amir Shafiq et al. introduced SalSi in SalSi: A new seismic attribute for salt dome detection, a novel seismic attribute that leverages principles of human visual attention to automatically flag prominent geological anomalies. Muhammad Amir Shafiq later expanded this concept in A novel attention model for salient structure detection in seismic volumes, combining 3D Fast Fourier Transform (3D-FFT) spectral decomposition with a directional center-surround attention model. This allows a single algorithm to detect multiple distinct feature types—such as faults and salt bodies—simultaneously by identifying where the visual "salience" of the data peaks.
Where traditional signal processing attributes struggle with noise, deep learning architectures provide robust alternatives. For instance, Zengyan Wang et al. designed a deep dilated attention autoencoder based on the U-Net architecture in Improved Automating Seismic Facies Analysis Using Deep Dilated Attention Autoencoders. By incorporating dilated convolutions, the network expands its receptive field to capture broader context without losing spatial resolution, while soft attention mechanisms force the model to focus on true structural boundaries rather than random noise. Even bio-inspired heuristics have found success; Walther Maciel et al. in Revisiting ant colony algorithms to seismic faults detection utilized refined Ant Colony Optimization (ACO) algorithms, where computational "ants" deposit pheromones along structural discontinuities, yielding sharp, automated maps of subtle geological faults.
Automating Waveform Analysis and Phase Picking
Beyond spatial volumes, the manual bottleneck is equally severe in continuous time-series data, where microseismic monitoring requires identifying the exact onset times of weak seismic phases.
Traditional automated pickers often fail when the signal-to-noise ratio (SNR) drops. To solve this, researchers have combined robust feature selection with clustering or multi-task deep networks. In First Arrival Picking on Microseismic Signals Based on K-Means with a ReliefF Algorithm, Yijia Li et al. paired the ReliefF algorithm for automated feature selection with K-Means clustering. This mathematical pairing filters out uninformative noise components before clustering, allowing the algorithm to pinpoint first arrival times consistently without human intervention.
[Continuous Waveform] ──► [Feature Selection (ReliefF)] ──► [K-Means Clustering] ──► Precise Arrival Pick
When deep learning is applied to waveforms, framing the task correctly is vital. Hang Zhang et al. developed PDNetwork, detailed in A Dual-Task Approach for Onset Time Picking and the Detection of Microseismic Waveforms Based on Deep Learning. Instead of treating waveform detection and onset picking as isolated steps, the model solves them simultaneously. By integrating a continuous "score map" that peaks precisely at the fracture event, the network leverages the contextual synergy of knowing that an event occurred to better determine when it started.
This strategy of reframing processing tasks into standard machine learning paradigms extends to velocity analysis. In Automatic Velocity Picking Using a Multi-Information Fusion Deep Semantic Segmentation Network, Hongtao Wang et al. reframed seismic velocity picking as a semantic segmentation task. Their Multi-Information Fusion Network (MIFN) blends raw velocity spectra with Stacked Gather Slices (SGS), allowing a deep network to "paint" the ideal velocity paths automatically, ensuring high-throughput analysis even in low-SNR environments.
Intelligent Denoising and Catalog Expansion
A major consequence of the manual bottleneck is that massive amounts of weak or noisy data are left unanalyzed, hiding critical geophysical insights. Automated workflows must therefore handle both noise isolation and signal amplification at scale.
Noise identification can be treated as an object detection problem. For example, Zhenghong Liang et al. adapted the YOLOv5 architecture in Airwave Noise Identification from Seismic Data Using YOLOv5 to automatically detect and isolate airwave noise profiles directly from seismic sections. Similarly, Matthias Meyer et al. addressed non-geophysical noise in Systematic identification of external influences in multi-year microseismic recordings using convolutional neural networks by combining seismic data with time-lapse camera imagery. Their CNN workflow successfully isolated human environmental noise (such as mountaineers on a rockface) from real tectonic signals, preventing human errors in long-term catalogs.
When applied to clean or intelligently denoised data, automated pipelines can dramatically expand our understanding of tectonic systems. In Micro-seismic events detection and its tectonic implications in Northeastern Hainan Province, the authors integrated deep-learning phase picking with GPU-accelerated template matching. By automating the search through years of continuous high-frequency data, they expanded the local earthquake catalogue from a mere 5 manually identified events to 977, mapping out a complex, active fault system that was previously invisible. Where machine learning labels are unavailable entirely, algorithmic automation still provides scalability; Darius Fenner et al. developed AWESAM (Automated Seismo-Volcanic Event Detection Applied to Stromboli (Italy)), utilizing an adaptive-window filtering technique to automatically catalog over 290,000 volcanic events without requiring any prior training labels.
Bypassing Labels with Unsupervised and Foundation Models
The latest evolution in overcoming the manual interpretation bottleneck moves away from supervised learning entirely. Supervised models simply shift the bottleneck from interpreting data to labeling data, which still requires expensive, error-prone human annotation. The frontier of this research direction relies on unsupervised learning and self-supervised foundation models.
To detect signals without human bias, Léonard Seydoux et al. developed an unsupervised deep learning framework detailed in Clustering earthquake signals and background noises in continuous seismic data with unsupervised deep learning. By pairing a learnable deep scattering network with Gaussian mixture models, the framework automatically clusters continuous, unlabeled data based on underlying statistical features. This system successfully uncovered subtle, unlabelled precursory signals prior to a massive landslide in Greenland—signals that standard supervised methods or human inspectors missed because they did not know what features to look for.
[Massive Unlabeled Waveform Data]
│
▼
[Self-Supervised Pretraining] ◄── Bypasses the human labeling bottleneck
│
▼
[Seismic Foundation Model]
│
┌────────┴────────┐
▼ ▼
[Phase Picking] [Facies Mapping] (Downstream tasks with minimal labels)
Taking this concept to its logical conclusion, Tianlin Liu et al. introduced SeisLM in SeisLM: a Foundation Model for Seismic Waveforms. Borrowing the self-supervised pretraining paradigms popularized by large language models, the authors trained an encoder-only Transformer on vast repositories of unlabeled seismic waveforms using contrastive learning.
The intuition behind SeisLM is powerful: by forcing the network to learn generalized representations of seismic wave propagation entirely on its own, the model develops an intrinsic "understanding" of the data. When deployed on downstream tasks like phase picking or event classification, it drastically outperforms traditional supervised models, especially when expert annotations are scarce. By shifting the paradigm from manual inspection to self-supervised foundation models, this research direction provides the mathematical and computational framework necessary to scale geophysical analysis to the petabyte era.
Key Foundational Work
SalSi: A new seismic attribute for salt dome detection (2016) — contributed 1 concept elements to this direction.
Tensor-based subspace learning for tracking salt-dome boundaries (2015) — contributed 1 concept elements to this direction.
Revisiting ant colony algorithms to seismic faults detection (2015) — contributed 1 concept elements to this direction.