Breccia and basalt classification of thin sections of Apollo rocks with deep learning
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.
Vera: We're looking at the paper: "Breccia and basalt classification of thin sections of Apollo rocks with deep learning".
Jocelyn: Here's the gist of the paper. Breccia and basalt classification of thin sections of Apollo rocks with deep learning introduces a framework leveraging contrastive learning to classify breccias from basalts in lunar rock thin sections with an accuracy of 98.44% The binary classifier of breccia/basalt rock textures with transfer learning from Inception-ResNet-v2 achieves an accuracy of 98.44% (±1.47) Human exploration of the moon is expected to resume in the next decade, following the last such activities in the Apollo programme time One of the major objectives of returning to the Moon is to continue retrieving geological samples, with a focus on collecting high-quality specimens to maximize scientific return A lunar rock classifier is a tool that can potentially provide the necessary information for astronauts to analyze lunar rock samples, allowing them to augment in-situ value identification of samples This study introduces a method for classifying rock textures in thin sections of lunar breccias and basalts and establishes a framework for additional analysis, including segmentation and clustering Data Sources and Preparation The study utilizes two databases, the NASA PDS database
25: and the Lunar Institute Data
26: , which contain digitally scanned versions of original film photographs of rock, soil, and core samples collected from the lunar surface during the Apollo missions The dataset comprises 365 samples with 5203 images of thin sections of the rock images Data preprocessing involved resizing images to 299x299x3 and normalizing pixel color values, followed by data augmentation including random vertical/horizontal rotations, random zoom, brightness, saturation, hue, and contrast Class frequencies were calculated to generate class weights to counteract potential problems arising from the imbalance in the dataset Model Architecture and Training Strategy The methodology employs contrastive learning to extract significant representations from images of thin sections of lunar rock The network consists of a transfer model of convolutional residual connections and a classification head The transfer learning model is an Inception-ResNet-v2 network pre-trained on ImageNet, which is then fine-tuned using the SimCLR loss function Contrastive learning involves augmenting input images into two representations, ˆxi and ˆxj, passing them through the neural network to get feature representations hi and hj The loss function maximizes agreement between the two representations while minimizing agreement between representations of other images The training is divided into two phases: first phase freezes the transfer learning model and trains the classification head, and second phase fine-tunes the combined transfer model and classification head Performance Metrics and Future Directions The resulting model proved successful, with an image accuracy of 93.51% (±3.17) and a group accuracy of 98.83% (±0.95), calculated on the test set The binary breccia/basalt classifier achieved an accuracy of 98.44% (±1.47) Future research is suggested to further categorize breccia into subgroups and include anorthosites in the analysis The transfer model, fine-tuned with the SimCLR loss function, can be utilized as a transfer model for other analyses using the same dataset In this study, we investigated the use of SimCLR, a contrastive learning method, for extracting features from lunar thin sections The binary breccia/basalt classifier achieved an accuracy of 98.44% (±1.47) The classification task in this paper is relatively straightforward, as there are 513 samples total, and the classification is binary between basalt and breccia The transfer model, fine-tuned with the SimCLR loss function, can be utilized as a transfer model for other analyses using the same dataset The expansion of the dataset would enhance the network’s effectiveness for other tasks Acknowledgements We extend our sincere acknowledgements to the co-authors Victoria S.
Vera: Engelschiøn, Piotr Knapczyk, Jonas Lewe, Clara Moriceau, and Romeo Haak Acknowledgements to NASA and LPI for providing the vast dataset of images of the thin sections of lunar rocks The research presented in this paper was done during their time as integral members of the Spaceship EAC group at the European Astronaut Centre, part of ESA References
1: B.
Jocelyn: L.
Vera: Jolliff, M.
Title and authors: Vera: Okay, so let’s start with the title and authors of "Breccia and basalt classification of thin sections of Apollo rocks with deep learning." It’s Freja Thoresen, Aidan Cowley, Romeo Haak, Jonas Lewe, Clara Moriceau, Piotr Knapczyk, Victoria S. Engelschiøn.
Jocelyn: That’s a solid team of researchers involved in this work. I wonder what the specific focus of that title tells us about the paper's scope.
Vera: The title is very direct: "Breccia and basalt classification of thin sections of Apollo rocks with deep learning." It clearly states the materials they are classifying—breccias and basalts—and exactly what they are using to analyze them, which is deep learning on thin section images.
Subrahmanyan: From a theoretical standpoint, I'm interested in how well the model captures the mineralogical differences between breccia and basalt at the scale of those thin sections.
Jocelyn: It sounds like they are trying to solve a real practical problem for planetary science—giving astronauts tools to analyze samples quickly instead of having to rely solely on painstaking optical microscopy.
Vera: Precisely, because the paper mentions that petrographic thin sections take time and expertise, and this method aims to create an automated or semi-automated system for analyzing lunar rocks.
Subrahmanyan: If you can automate the textural identification of these key rock types, it streamlines geological interpretation for understanding lunar formation processes.
Jocelyn: I think that’s the key implication here: moving from slow manual analysis to a fast classification tool that could be used in real-time by explorers.
Vera: That’s what they are aiming for with the overall goal of returning to the Moon and maximizing scientific return from sample collection.
Subrahmanyan: It connects directly to our work on structure formation; being able to quickly categorize these components helps us map out the geological history of the lunar crust better.
Jocelyn: So, it’s bridging the gap between high-resolution microscopy and automated, scalable data processing for space exploration.
Vera: That’s a good way to put it—it connects high-resolution imaging with scalable data processing for future missions.
The paper's summary: Vera: Now let’s look at the summary of "Breccia and basalt classification of thin sections of Apollo rocks with deep learning." They say they are applying state-of-the-art machine learning techniques to a comprehensive dataset of lunar rock thin section images.
Jocelyn: And they specifically highlight contrastive learning as their method for extracting significant features from those images, which is quite interesting.
Vera: Yes, the paper emphasizes that contrastive learning is used to capture the underlying features of those images, rather than just relying on superficial patterns that a standard classifier might pick up.
Subrahmanyan: That’s important because it suggests the method is learning more abstract, physically meaningful representations of what makes a breccia fundamentally different from a basalt.
Jocelyn: And then they build on top of that using transfer learning from an Inception-ResNet-v2 network to create a binary classifier for the breccia/basalt rock textures.
Vera: So, they freeze the transfer model initially and train only the classification head, and then they fine-tune everything together in a second phase. This two-step training is part of their methodology.
Subrahmanyan: It’s a standard but effective way to manage complexity; it lets them get initial feature representations stable before optimizing the final decision boundary for the specific task.
Jocelyn: And what’s the practical outcome of all this? They report an accuracy of ninety-eight point four four plus or minus one point four seven for their binary breccia/basalt classifier on the test set.
Vera: That high accuracy is what they claim, and it shows that their method performs very well in distinguishing between those two rock types given the data they used.
Subrahmanyan: It’s a strong result, but we always look at those error bars; the one point four seven uncertainty tells us how robust this classification is when applied to new, unseen lunar samples.
The paper's improvements: Vera: The paper also points out some specific improvements they suggest for future work. They mention expanding the analysis to include anorthosites and further categorizing breccias into subgroups.
Jocelyn: So, they’re looking beyond just a simple binary split of breccia versus basalt; they want a more detailed geological breakdown of the breccias themselves.
Vera: Right, because if you can categorize the breccias into subgroups, you move from just saying "this is a breccia" to understanding *what kind* of breccia it is.
Subrahmanyan: That level of detail is exactly what we need for structure formation studies; knowing the specific composition and texture variations helps us constrain models for how lunar material accreted over time.
Jocelyn: And another key improvement mentioned is that they suggest the transfer model can be utilized as a transfer model for other analyses using the same dataset.
Vera: That means if you want to do segmentation or clustering on these images, you don't have to retrain the entire feature extraction network from scratch; you can just take the fine-tuned model and apply it.
Subrahmanyan: That makes sense because training those large feature extractors is computationally expensive, so reusing that pre-trained knowledge for other tasks is a major efficiency gain.
Jocelyn: It essentially makes this entire deep learning pipeline a versatile resource for analyzing lunar rock data, opening up avenues for more complex research.
Conclusion: Vera: So, to wrap up the paper "Breccia and basalt classification of thin sections of Apollo rocks with deep learning," they successfully demonstrated a binary breccia/basalt classifier reaching an accuracy of ninety-eight point four four plus or minus one point four seven.
Jocelyn: And they also pointed out that this transfer model can be used for other analyses, meaning the feature extraction network is a versatile resource for segmentation and clustering tasks on lunar rock data.
Subrahmanyan: For me, the biggest implication is that we’re getting a robust tool to rapidly categorize these components of lunar crust, which could significantly speed up geological interpretation.
Vera: It sets a clear path for how deep learning can be used to augment the analysis of Apollo rocks for future missions by providing fast, automated classification capabilities.
Jocelyn: It’s a very promising result that shows how powerful this combination of contrastive learning and transfer learning can be when applied to complex geological imaging tasks.
Subrahmanyan: I just hope that with more data, we can use these techniques to look at even finer details in the texture and get closer to understanding the formation processes of lunar rocks.
Freja Thoresen, Aidan Cowley, Romeo Haak, Jonas Lewe, Clara Moriceau, Piotr Knapczyk, Victoria S. Engelschiün
Directorate of Human and Robotics Exploration, European Space Agency (ESA) · European Space Research and Technology Centre (ESTEC) · The Natural History Museum, University of Oslo
astro-ph.EP, astro-ph.IM, cs.LG
Submitted: 2024-10-28
Updated: 2026-10-27
Comments: 6 pages, 5 figures, 2 tables. Accepted at the ESA SPAICE 2026 Conference (Space AI and Computing for Earth and Space). Code: https://github.com/esa/apollo_rock_thin_section_classifier
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Breccia and basalt classification of thin sections of Apollo rocks with deep learning introduces a framework leveraging contrastive learning to classify breccias from basalts in lunar rock thin…
Key concepts
- Contrastive Learning
- A machine learning technique where the model learns to understand data by pulling similar items closer together in a feature space while pushing dissimilar items further apart. In this study, it helps the network extract meaningful features from lunar rock images by comparing different views of the same sample.
- Inception-ResNet-v2
- A powerful deep learning model pre-trained on a massive dataset called ImageNet. This model is used as a transfer learning base, meaning its knowledge about general image features is adapted and fine-tuned specifically for classifying lunar rock textures, making the classification task much more effective.
- Transfer Learning
- A method where a model trained on one task (like identifying general images) is repurposed for a different, related task (like classifying lunar breccias). This saves time and data because the model already possesses strong foundational knowledge about visual patterns from its initial training.
- Binary Classification
- A type of machine learning problem where the goal is to sort data into exactly two distinct categories. In this research, the system is specifically designed to distinguish between two types of lunar rock thin sections: breccia and basalt.
Terminology
Summary
Breccia and basalt classification of thin sections of Apollo rocks with deep learning introduces a framework leveraging contrastive learning to classify breccias from basalts in lunar rock thin sections with an accuracy of 98. p. 2 Deep learning, especially convolutional neural networks (CNN), has impressive performance in
image classification and segmentation tasks [8, 9]. This study introduces a method for classifying rock textures in thin sections of lunar breccias and basalts and establishes a framework for additional
analysis, including segmentation and clustering. • Contrastive learning on lunar rock thin sections to capture the underlying features of the images. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98.Where the paper says it
Where the paper says it
Introduction to the Problem
Human exploration of the moon is expected to resume in the next decade, following the last such activities in the Apollo programme time One of the major objectives of returning to the Moon is to continue retrieving geological samples, with a focus on collecting high-quality specimens to maximize scientific return A lunar rock classifier is a tool that can potentially provide the necessary information for astronauts to analyze lunar rock samples, allowing them to augment in-situ value identification of samples This study introduces a method for classifying rock textures in thin sections of lunar breccias and basalts and establishes a framework for additional analysis, including segmentation and clustering
Data Sources and Preparation
The study utilizes two databases, the NASA PDS database [25] and the Lunar Institute Data [26], which contain digitally scanned versions of original film photographs of rock, soil, and core samples collected from the lunar surface during the Apollo missions The dataset comprises 365 samples with 5203 images of thin sections of the rock images Data preprocessing involved resizing images to 299x299x3 and normalizing pixel color values, followed by data augmentation including random vertical/horizontal rotations, random zoom, brightness, saturation, hue, and contrast Class frequencies were calculated to generate class weights to counteract potential problems arising from the imbalance in the dataset
Model Architecture and Training Strategy
The methodology employs contrastive learning to extract significant representations from images of thin sections of lunar rock The network consists of a transfer model of convolutional residual connections and a classification head The transfer learning model is an Inception-ResNet-v2 network pre-trained on ImageNet, which is then fine-tuned using the SimCLR loss function Contrastive learning involves augmenting input images into two representations, ˆxi and ˆxj, passing them through the neural network to get feature representations hi and hj The loss function maximizes agreement between the two representations while minimizing agreement between representations of other images The training is divided into two phases: first phase freezes the transfer learning model and trains the classification head, and second phase fine-tunes the combined transfer model and classification head
Performance Metrics and Future Directions
The resulting model proved successful, with an image accuracy of 93.51% (±3.17) and a group accuracy of 98. p. 7 Figure 4 shows the structure of the transfer learning model. The first 6
blocks depicted are a group of multiple layers (mostly convolutional) defined
in [31]. • Contrastive learning on lunar rock thin sections to capture the underlying features of the images. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98.Where the paper says it
Where the paper says it
Conclusion
In this study, we investigated the use of SimCLR, a contrastive learning method, for extracting features from lunar thin sections The binary breccia/basalt classifier achieved an accuracy of 98.44% (±1.47) The classification task in this paper is relatively straightforward, as there are 513 samples total, and the classification is binary between basalt and breccia The transfer model, fine-tuned with the SimCLR loss function, can be utilized as a transfer model for other analyses using the same dataset The expansion of the dataset would enhance the network’s effectiveness for other tasks
Acknowledgements
We extend our sincere acknowledgements to the co-authors Victoria S. Engelschiøn, Piotr Knapczyk, Jonas Lewe, Clara Moriceau, and Romeo Haak Acknowledgements to NASA and LPI for providing the vast dataset of images of the thin sections of lunar rocks The research presented in this paper was done during their time as integral members of the Spaceship EAC group at the European Astronaut Centre, part of ESA
References
[1] B. L. Jolliff, M. A. Wieczorek, C. K. Shearer, C. R. Neal, New views of the Moon, volume 60, Mineralogical Society of America Chantilly, 2006 [2] I. p. 3 In Figure 1, a concept of a semi-automatic machine is
shown, which was developed in 1965. Development of tools and payloads
to support astronauts with in-situ science analysis capability represents a
new paradigm for exploration and will be increasingly important in future
missions beyond LEO - ultimately, a level of laboratory capability will be
required as Moon and even Mars missions make terrestrial return logistics
challenging. Engelschiønb,c
aDirectorate of Human and Robotics Exploration, European Space Agency (ESA),
European Space Research and Technology Centre (ESTEC), Noordwijk, Netherlands
bDirectorate of Human and Robotics Exploration, European Space Agency (ESA),
European Astronaut Centre (EAC), Cologne, Germany
cThe Natural History Museum, University of Oslo, Oslo, Norway
Abstract
Human exploration of the moon is expected to resume in the next decade,
following the last such activities in the Apollo programme time. One of the
major objectives of returning to the Moon is to continue retrieving geological
samples, with a focus on collecting high-quality specimens to maximize scientific return. This information can help guide further
collection efforts, allowing targeted sampling of promising sites and maximizing the scientific returns of lunar missions for analysis back on Earth,
in addition to providing impetus for the development of IVA (in-vehicle
activity) thin-section/microscopy capabilities for future lunar astronauts. Building Blocks of the Moon
The Moon mainly consists of the igneous rocks anorthosite and basalt,
and breccias, which are aggregates of broken rock fragments cemented together. One of the
major objectives of returning to the Moon is to continue retrieving geological
samples, with a focus on collecting high-quality specimens to maximize scientific return. Tools that assist astronauts in making informed decisions about
sample collection activities can maximize the scientific value of future lunar
missions. Data
Two databases are used in this study: The NASA PDS database [25] and
the Lunar Institute Data [26]. The NASA PDS database contains digitally
scanned versions of the original film photographs of rock, soil, and core
samples collected from the lunar surface during the Apollo missions. Breccia and basalt classification of thin sections of Apollo
rocks with deep learning
Freja Thoresena,, Aidan Cowleyb
, Romeo Haakb
, Jonas Leweb
, Clara
Moriceaub
, Piotr Knapczykb
, Victoria S. Engelschiønb,c
aDirectorate of Human and Robotics Exploration, European Space Agency (ESA),
European Space Research and Technology Centre (ESTEC), Noordwijk, Netherlands
bDirectorate of Human and Robotics Exploration, European Space Agency (ESA),
European Astronaut Centre (EAC), Cologne, Germany
cThe Natural History Museum, University of Oslo, Oslo, Norway
Abstract
Human exploration of the moon is expected to resume in the next decade,
following the last such activities in the Apollo programme time. The approach segments petrographic thin
section components in tight oil reservoirs, demonstrating improvements in
accuracy and speed [22]. In [23], a method for classifying thin-section images
using residual networks (ResNet50 and ResNet101) is introduced in order to
achieve automatic feature extraction and classification based on image size. Introduction
The analysis of lunar samples is crucial to understanding the formation,
evolution, and resource potential of the Moon [1]. Hence, rock sample collection is one of the important scientific objectives of returning to the Moon. In recent years, machine learning has shown to be a powerful tool
for automating the analysis of thin-section images [7]. Deep learning, especially convolutional neural networks (CNN), has impressive performance in
image classification and segmentation tasks [8, 9]. Building Blocks of the Moon
The Moon mainly consists of the igneous rocks anorthosite and basalt,
and breccias, which are aggregates of broken rock fragments cemented together. Anorthosite is a calcium-rich rock that largely makes up the ancient
crust of the Moon [10]. This information can help guide further
collection efforts, allowing targeted sampling of promising sites and maximizing the scientific returns of lunar missions for analysis back on Earth,
in addition to providing impetus for the development of IVA (in-vehicle
activity) thin-section/microscopy capabilities for future lunar astronauts. Building Blocks of the Moon
The Moon mainly consists of the igneous rocks anorthosite and basalt,
and breccias, which are aggregates of broken rock fragments cemented together. Segmentation tasks focus on classifying individual pixels within an image, while classification approaches aim to categorize entire images based on features like mineral composition or grain size. Data
Two databases are used in this study: The NASA PDS database [25] and
the Lunar Institute Data [26]. Figure 1: In 1965, the United States Geological Survey (USGS) developed a prototype for
a semi-automatic machine to cut and prepare thin sections of rocks [20]. This innovative
device was specifically designed for the Apollo Applications Program and Advanced Lunar
Programs Groups, aiming to streamline the analysis of lunar samples. In Figure 1, a concept of a semi-automatic machine is
shown, which was developed in 1965. Development of tools and payloads
to support astronauts with in-situ science analysis capability represents a
new paradigm for exploration and will be increasingly important in future
missions beyond LEO - ultimately, a level of laboratory capability will be
required as Moon and even Mars missions make terrestrial return logistics
challenging. Development of tools and payloads
to support astronauts with in-situ science analysis capability represents a
new paradigm for exploration and will be increasingly important in future
missions beyond LEO - ultimately, a level of laboratory capability will be
required as Moon and even Mars missions make terrestrial return logistics
challenging. Analysis of thin sections of lunar rock using machine learning techniques
Previous studies have utilized machine learning techniques to cluster,
classify, or segment thin sections of rocks. Development of tools and payloads
to support astronauts with in-situ science analysis capability represents a
new paradigm for exploration and will be increasingly important in future
missions beyond LEO - ultimately, a level of laboratory capability will be
required as Moon and even Mars missions make terrestrial return logistics
challenging. Analysis of thin sections of lunar rock using machine learning techniques
Previous studies have utilized machine learning techniques to cluster,
classify, or segment thin sections of rocks. Development of tools and payloads
to support astronauts with in-situ science analysis capability represents a
new paradigm for exploration and will be increasingly important in future
missions beyond LEO - ultimately, a level of laboratory capability will be
required as Moon and even Mars missions make terrestrial return logistics
challenging. Analysis of thin sections of lunar rock using machine learning techniques
Previous studies have utilized machine learning techniques to cluster,
classify, or segment thin sections of rocks. Towards demonstrating the value of such a tool, in this paper, we introduce a framework
for classifying rock types in thin sections of lunar rocks. We leverage the
vast collection of petrographic thin-section images from the Apollo missions,
captured under plane-polarized light (PPL), cross-polarised light (XPL),
and reflected light at varying magnifications. Towards demonstrating the value of such a tool, in this paper, we introduce a framework
for classifying rock types in thin sections of lunar rocks. We leverage the
vast collection of petrographic thin-section images from the Apollo missions,
captured under plane-polarized light (PPL), cross-polarised light (XPL),
and reflected light at varying magnifications. Development of tools and payloads
to support astronauts with in-situ science analysis capability represents a
new paradigm for exploration and will be increasingly important in future
missions beyond LEO - ultimately, a level of laboratory capability will be
required as Moon and even Mars missions make terrestrial return logistics
challenging. Analysis of thin sections of lunar rock using machine learning techniques
Previous studies have utilized machine learning techniques to cluster,
classify, or segment thin sections of rocks. While thin sections of lunar rocks were created when they were returned
and analyzed on Earth, an interesting prospect is to make the thin sections
in a lunar laboratory. In [19], designs for an automated rock thin section
device are suggested, which may efficiently cut future rock samples in a
space environment. Figure 1: In 1965, the United States Geological Survey (USGS) developed a prototype for
a semi-automatic machine to cut and prepare thin sections of rocks [20]. This innovative
device was specifically designed for the Apollo Applications Program and Advanced Lunar
Programs Groups, aiming to streamline the analysis of lunar samples. In [19], designs for an automated rock thin section
device are suggested, which may efficiently cut future rock samples in a
space environment. In Figure 1, a concept of a semi-automatic machine is
shown, which was developed in 1965. Development of tools and payloads
to support astronauts with in-situ science analysis capability represents a
new paradigm for exploration and will be increasingly important in future
missions beyond LEO - ultimately, a level of laboratory capability will be
required as Moon and even Mars missions make terrestrial return logistics
challenging. Analysis of thin sections of lunar rock using machine learning techniques
Previous studies have utilized machine learning techniques to cluster,
classify, or segment thin sections of rocks. The approach segments petrographic thin
section components in tight oil reservoirs, demonstrating improvements in
accuracy and speed [22]. In [23], a method for classifying thin-section images
using residual networks (ResNet50 and ResNet101) is introduced in order to
achieve automatic feature extraction and classification based on image size. In [23], a method for classifying thin-section images
using residual networks (ResNet50 and ResNet101) is introduced in order to
achieve automatic feature extraction and classification based on image size. Experimental results show that the two network structures achieve high accuracy (90. Another approach uses CNNs to classify 13 rock types [24]. A concatenated Convolutional Neural Network is used to classify rocks based on
photographed thin sections in PPL and XPL lighting conditions. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98. 44
arXiv:2410. There are 365 samples in the dataset, with 5203 images of thin
sections of the rock images, as one rock sample can have several images. Like the NASA PDS database, the Lunar Institute Data also contains
digitally scanned versions of original film photographs of rock, soil, and core
samples collected from the Apollo missions. There are 365 samples in the dataset, with 5203 images of thin
sections of the rock images, as one rock sample can have several images. Like the NASA PDS database, the Lunar Institute Data also contains
digitally scanned versions of original film photographs of rock, soil, and core
samples collected from the Apollo missions. Further details on the samples were retrieved from the
Lunar Sample and Photo Catalog Curator [27] to fill in gaps of knowledge
where needed. Images with mislabelling, manual markings, or where the thin
section occupied less than 60% of the image area were excluded from the
dataset. Highlights
Breccia and basalt classification of thin sections of Apollo rocks
with deep learning
Freja Thoresen, Aidan Cowley, Romeo Haak, Jonas Lewe, Clara Moriceau,
Piotr Knapczyk, Victoria S. Engelschiøn
• State-of-the-art machine learning techniques and analyses applied to
a comprehensive dataset of lunar rock thin section images. In recent years, machine learning has shown to be a powerful tool
for automating the analysis of thin-section images [7]. Deep learning, especially convolutional neural networks (CNN), has impressive performance in
image classification and segmentation tasks [8, 9]. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98. 44
arXiv:2410. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98. 44
arXiv:2410. ResNet is a convolutional neural architecture that vastly improved the accuracy of deep neural
networks by implementing residual connections. It was found that deeper convolutional networks with the inception principles proved effective in improving the use of computing resources within
the network, hence improving the network while keeping computational resources constant [30]. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98. 44
arXiv:2410. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98. 44
arXiv:2410. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98. 44
arXiv:2410. In Figure 1, a concept of a semi-automatic machine is
shown, which was developed in 1965. Development of tools and payloads
to support astronauts with in-situ science analysis capability represents a
new paradigm for exploration and will be increasingly important in future
missions beyond LEO - ultimately, a level of laboratory capability will be
required as Moon and even Mars missions make terrestrial return logistics
challenging. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98. 44
arXiv:2410. • Binary classifier of breccia/basalt rock textures with transfer learning
from Inception-ResNet-v2 achieves an accuracy of 98. 44
arXiv:2410.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Contrastive learning on lunar rock thin sections to capture underlying features of images allows for fine-tuned feature extraction from Inception-ResNet-v2, which can then be used as a transfer model for other analyses like segmentation and clustering.
-
A binary classifier trained using transfer learning from the fine-tuned Inception-ResNet-v2 achieves an accuracy of 98.44% (±1.47) in separating breccias from basalts, enabling automated classification of lunar rock thin sections for sample identification.
-
The system can be utilized as a transfer model for other analyses using the same dataset, opening the possibility of continuing research by employing the feature extraction network for segmentation and clustering approaches.
Abstract
Human exploration of the moon is expected to resume in the next decade, following the last such activities in the Apollo programme time. One of the major objectives of returning to the Moon is to continue retrieving geological samples, with a focus on collecting high-quality specimens to maximize scientific return. Tools that assist astronauts in making informed decisions about sample collection activities can maximize the scientific value of future lunar missions. A lunar rock classifier is a tool that can potentially provide the necessary information for astronauts to analyze lunar rock samples, allowing them to augment in-situ value identification of samples. Towards demonstrating the value of such a tool, in this paper, we introduce a framework for classifying rock types in thin sections of lunar rocks. We leverage the vast collection of petrographic thin-section images from the Apollo missions, captured under plane-polarized light (PPL), cross-polarised light (XPL), and reflected light at varying magnifications. Advanced machine learning methods, including contrastive learning, are applied to analyze these images and extract meaningful features. The contrastive learning approach fine-tunes a pre-trained Inception-Resnet-v2 network with the SimCLR loss function. The fine-tuned Inception-Resnet-v2 network can then extract essential features effectively from the thin-section images of Apollo rocks. A simple binary classifier is trained using transfer learning from the fine-tuned Inception-ResNet-v2 to 98.44% (plus or minus 1.47) accuracy in separating breccias from basalts.
Sources
- Evolution of Image Segmentation using Deep Convolutional Neural Network: A Survey
- Rock Classification in Petrographic Thin Section Images Based on Concatenated Convolutional Neural Networks
- The Effectiveness of Data Augmentation in Image Classification using Deep Learning
- Deep Residual Learning for Image Recognition
- Rethinking the Inception Architecture for Computer Vision
- Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
- A Simple Framework for Contrastive Learning of Visual Representations
- Adam: A Method for Stochastic Optimization
- On the Convergence of Adam and Beyond
Related papers
- A Bright L4 Neptune Trojan Found by Stacking Archival ZTF Images
- PDS 70 c and SR 12 c: Observational Constraints on Giant-Planet and Satellite Formation
- JWST Rocky Worlds DDT Observations of LTT 1445 A b and c in Secondary Eclipse Using MIRI/F1500W
- A Joint Study of Epoch Astrometry and Radial Velocities of Gaia-4b
- Using Structure Functions to Characterize Protoplanetary Disk Velocity Residuals
- An analytical framework for atmospheric tides on rocky planets. II. Semidiurnal thermotidal torque