📄 earth_science

MTA-Net for Predicting Grain-Size Composition Distributions from Sedimentary Simulation Experiment Images by Integrating Multiscale Texture Features

This paper introduces MTA-Net, a deep learning model that integrates ResNet34 with Horizontal Multi-Scale Local Binary Pattern (HMS-LBP) and Convolutional Block Attention Module (CBAM) to accurately predict grain-size composition distributions from sedimentary experiment images by effectively capturing multiscale texture features and overcoming challenges like particle overlap.

Original authors: Yanlin Shao, Jiaxuan Wang, Wei Wei, Guangming Hu

Published 2026-07-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yanlin Shao, Jiaxuan Wang, Wei Wei, Guangming Hu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of looking for fingerprints, you are looking at a pile of sand. In the world of geology, scientists often need to know exactly how big the grains in a pile of sand are. This "grain-size" tells a story about how the sand got there—whether it was carried by a roaring river, a gentle breeze, or a crashing wave. Usually, to find this out, geologists have to take a physical scoop of sand, dry it out, and run it through a machine in a lab. It's accurate, but it's slow, messy, and you can't do it continuously while watching a river change in real-time.

Recently, scientists have started trying to use cameras instead of scoops. They take pictures of the sand and hope a computer can "see" the grain sizes just by looking at the texture, the patterns, and how the light hits the bumps and dips. It's a bit like trying to guess the size of pebbles on a beach just by squinting at a photograph from a distance. The problem is that sand grains often overlap, hide behind each other, and create a messy, jumbled texture that is hard for computers to decode. If the computer gets it wrong, the whole story about how the sand was deposited gets twisted.

This is where a new study comes in, acting like a super-powered magnifying glass for computer vision. A team of researchers has built a smart computer program called MTA-Net (Multi-scale Texture Attention Network) that is designed specifically to look at photos of sediment and guess the mix of grain sizes inside. Think of it as teaching a robot to not just "see" the sand, but to "feel" the texture of the sand through its eyes.

Here is how they did it and what they found:

The Detective's Toolkit: Mixing Old Tricks with New Brains
The researchers started with a standard, powerful computer brain known as ResNet34. You can think of this as a very smart student who is good at recognizing general shapes and patterns. However, when it comes to the messy, overlapping world of sand, this student sometimes misses the subtle clues.

To help this student, the researchers added two special tools:

HMS-LBP (Horizontal Multi-Scale Local Binary Pattern): Imagine this as a special pair of glasses that only looks at the sand horizontally. Since sand often settles in layers (like a cake), this tool scans the image left-to-right to catch the "stripes" and patterns that regular cameras might miss. It's like a detective who knows that footprints in the mud usually follow a straight line, so they only look for lines, not circles.
CBAM (Convolutional Block Attention Module): This is like a highlighter pen. Once the student has looked at the whole picture, this tool tells them, "Hey, look here! The important clues are in this specific spot, and ignore the boring background." It helps the computer focus on the parts of the image that actually tell the story of the grain sizes.

The Big Test: Can the Robot Read the Sand?
The team tested their new MTA-Net system using photos from a controlled lab experiment where they simulated how sand settles in water. They had 202 photos, and for each one, they knew the exact answer because they had physically measured the sand in the lab.

The results were impressive. When they compared their new MTA-Net to the standard ResNet34 student:

The new system made 45.41% fewer mistakes in guessing the overall distribution of grain sizes.
It reduced the average error in guessing specific grain sizes by 29.71%.
It got much closer to the "true" answer, with a score called NRMSE dropping to 16.53% (compared to 18.92% for the standard model).

In plain English, the MTA-Net was significantly better at reading the "texture story" of the sand than the standard computer models.

What the Computer Learned (and What It Didn't)
The researchers also ran a series of "what if" experiments to see which part of the toolkit was doing the heavy lifting.

They found that just adding the "horizontal glasses" (HMS-LBP) helped a lot.
They found that just adding the "highlighter" (CBAM) helped a bit, but not as much.
Crucially, they found that having both together was the magic combination. It wasn't just that the computer got bigger or smarter; it was that the two tools worked together perfectly. One helped the computer understand the structure of the sand layers, and the other helped it focus on the right details.

They also tested if the direction of the "glasses" mattered. They tried glasses that looked in all directions, or only vertically, but the horizontal ones worked best. This suggests that the way sand settles in their experiments creates horizontal patterns that are key to solving the puzzle.

The Limits of the Magic
While the robot is great, it's not a magician. The study showed that the system works best when the photos are clear and high-resolution. If you blur the photo or zoom out too far (downsampling), the robot gets confused because the tiny texture clues disappear.

Also, the robot struggled a bit with the very smallest grains (less than 100 micrometers) and the very largest ones (over 1000 micrometers). This is likely because the small grains are hard to see in the mix, and the large grains are so few and far between that they don't create a consistent pattern. The system is best at guessing the "middle" sizes, which make up the bulk of the sand.

The Bottom Line
This paper doesn't claim to have solved the problem of analyzing sand forever. Instead, it suggests that by combining a standard computer brain with specific, human-designed tools that understand how sand textures work (like horizontal layers), we can get much better at predicting grain sizes from photos. It's a step toward letting geologists watch sedimentary experiments in real-time, using a camera and a smart algorithm instead of a shovel and a lab bench. The method is promising, but the researchers admit it needs to be tested on different types of sand and different lighting conditions before it can be used everywhere.

Technical Summary: MTA-Net for Predicting Grain-Size Composition Distributions

Problem Statement
Laboratory sedimentation experiments are essential for investigating sediment transport, deposition, and sorting under controlled conditions. However, conventional grain-size analysis relies on physical sampling and laboratory testing (e.g., sieving, laser diffraction), which limits the ability to perform rapid, spatially continuous characterization of experimental deposits. While image-based analysis offers a non-contact alternative, it faces significant challenges: sediment images often exhibit particle overlap, indistinct boundaries, and complex, intertwined multiscale textures. Furthermore, existing segmentation-based methods struggle with occlusion, and traditional texture-based methods using handcrafted features often fail to model the nonlinear relationship between complex image textures and multicomponent grain-size distributions. The core challenge is to integrate sedimentologically meaningful texture statistics into deep learning feature learning to predict a five-component grain-size composition vector from a single image.

Methodology
The authors propose the Multi-scale Texture Attention Network (MTA-Net), a deep learning architecture designed to estimate grain-size composition distributions from sedimentation simulation images. The methodology integrates explicit texture descriptors with convolutional feature learning through a hierarchical design:

Backbone and Architecture: MTA-Net utilizes ResNet34 as its backbone for processing RGB images. The network is structured to inject texture guidance at the mid-level and apply attention refinement at the high-level.
HMS-LBP Texture Branch: A key innovation is the Horizontal Multi-Scale Local Binary Pattern (HMS-LBP). Unlike standard LBP, HMS-LBP is designed to capture direction-sensitive, multiscale structures common in sediment profiles (e.g., bedding-like horizontal continuity).
- It defines a horizontally enhanced neighborhood for a center pixel, emphasizing horizontal continuity while providing vertical context.
- It operates at multiple radii ( $r=\{1, 2\}$ ) to capture both local particle details and broader structural patterns.
- The resulting descriptors are extracted from grayscale images and fused via average pooling to create a 944-dimensional vector.
Mid-Level Texture Injection: The HMS-LBP descriptor is not merely concatenated at the output. Instead, it is injected after Stage 2 of the ResNet34 backbone. The texture vector is projected and used to modulate the channel features of the Stage 2 feature map via a learnable scaling coefficient. This allows explicit texture statistics to guide the formation of mid-level structural representations.
High-Level Attention: After Stage 4, the network applies the Convolutional Block Attention Module (CBAM). This refines the high-level features by generating channel and spatial attention maps, enhancing responses related to grain-size distribution while suppressing irrelevant background noise.
Loss Function: The model employs a composite distribution loss combining Cumulative Distribution Function (CDF), L1, and cosine losses. Crucially, it incorporates an entropy-based weighting scheme to place greater emphasis on training samples with dispersed compositions (higher entropy), addressing the difficulty of predicting transitional samples.

Key Contributions

Novel Architecture: The development of MTA-Net, which uniquely combines a ResNet backbone with mid-level HMS-LBP texture injection and high-level CBAM attention.
Direction-Sensitive Texture Encoding: The introduction of HMS-LBP, which explicitly models the horizontal and multiscale nature of sedimentary textures, outperforming standard, rotation-invariant, and vertically enhanced LBP variants.
Strategic Module Placement: Demonstrating through ablation studies that injecting texture guidance at Stage 2 and applying attention at Stage 4 yields superior performance compared to alternative insertion points or using modules in isolation.
Compositional Regression: Successfully framing the task as a constrained regression problem to predict a five-component volume fraction vector, rather than simple classification or single-value estimation.

Results
The model was evaluated on a dataset of 202 image-label pairs derived from a laboratory sedimentation experiment simulating Jurassic deltaic deposits. The data was split into training (145), validation (26), and test (31) sets.

Performance Metrics: MTA-Net achieved a $\chi^2$ distance of 0.0125 ± 0.0023, a Mean Absolute Error (MAE) of 0.0310 ± 0.0024, and a Normalized Root Mean Square Error (NRMSE) of 16.53 ± 0.36%.
Comparative Improvement: Compared to the baseline ResNet34, MTA-Net reduced errors by 45.41% ( $\chi^2$ ), 29.71% (MAE), and 12.63% (NRMSE). It also outperformed other baselines including ResNet50, EfficientNet-B0, and traditional machine learning approaches (Wavelet+RF, GLCM+RF, LBP+RF).
Ablation Findings:
- Adding HMS-LBP alone provided significant gains over standard LBP, confirming the value of directional and multiscale encoding.
- CBAM alone improved performance but less than HMS-LBP, indicating that attention refinement complements but does not replace explicit texture guidance.
- The combination of both modules yielded the best results.
- Sensitivity analysis confirmed that the $r=\{1, 2\}$ radius configuration and the specific insertion stages (Stage 2 for HMS-LBP, Stage 4 for CBAM) were optimal.
Limitations Observed: The model showed higher error rates for fine fractions (<100 μm and 100–250 μm) which were rarely dominant in the dataset, and performance degraded significantly when input images were downsampled by factors of four or more, highlighting the dependency on sufficient texture resolution.

Significance and Claims
The paper claims that MTA-Net provides a feasible texture-guided approach for estimating grain-size composition from sedimentation images. The authors emphasize that the success of the model stems not merely from increased network depth or size, but from the complementary contributions of explicit, sedimentologically meaningful texture statistics (HMS-LBP) and learned discriminative features (CBAM).

The study concludes that while the method is effective under the specific controlled experimental and imaging conditions used, its broader application to natural settings or different imaging systems requires further validation. The authors modestly position the work as a step toward quantitative analysis of sedimentation experiments, noting that future work must address cross-condition transferability, robustness to illumination variations, and the need for larger, multisource datasets before widespread sedimentological application.

More like this