astronomy/ai produced the result/Nature Communications 2023 · v2
Neural network turns Kaguya camera images into global maps of lunar surface chemistry
Researchers trained a neural network on laboratory measurements of 115 lunar soil samples, including Chang'e-5 material, then used it to predict the abundances of six major oxides across the Moon from orbital camera data.
spectrum · one line per step, placed by what the step does · bright lines used AI
Comprehensive mapping of lunar surface chemistry by adding Chang'e-5 samples with deep learning
Nature Communications, 2023
doi:10.1038/s41467-023-43358-0 · record aix-00209 v2 · checked 2026-10-09
- AI was for
- Property prediction
- Model family
- Convolutional neural network, Multilayer perceptron
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
The Moon's surface rock is made of a handful of common chemical ingredients: compounds of titanium, iron, aluminium, magnesium, calcium and silicon, usually reported as oxides. Knowing how much of each lies where tells geologists which lavas erupted when, and how the crust formed. The trouble is that only a few dozen spots have ever been sampled directly, by the Apollo and Luna missions and, more recently, by China's Chang'e-5. Everywhere else, chemistry has to be inferred from orbit, by reading the sunlight that the surface reflects. Different minerals reflect slightly different amounts of light at different wavelengths, but the link between a reflectance spectrum and an actual chemical recipe is indirect and messy.
The researchers set out to build that link from the ground truth available. They paired laboratory-measured oxide abundances from 115 lunar soil samples, collected at 55 sampling sites across the Apollo, Luna and Chang'e-5 landing regions, with eight-wavelength reflectance measurements of the same places taken by the Multiband Imager on Japan's SELENE (Kaguya) orbiter. The aim was a model that could then be applied to every pixel of the imagery, producing global maps of the six oxides.
Where AI came in
The model is a one-dimensional convolutional neural network, a kind of network suited to reading a short sequence of numbers — here, the eight brightness values of a single pixel, plus two extra quantities calculated to reduce the effect of long exposure to space weathering. It was trained from scratch to output the six oxide abundances. Training used a leave-one-out scheme, in which the model is retrained many times with one sample held back each round to check its predictions; the resulting models were then merged by averaging their internal weights into a single predictor.
That predictor supplied every value in the resulting maps, standing in for the statistical fitting formulas previously used to convert reflectance into chemistry. The team also trained simpler machine-learning models for comparison, and built a second version of their network without the Chang'e-5 samples. The finished maps were then checked against measurements made on the surface by the Chang'e-3 and Chang'e-4 missions and against earlier published chemistry maps, before being used by the researchers themselves to draw boundaries between lunar geological regions and to define 26 compositional units among the younger lava plains.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
A one-dimensional convolutional neural network was trained to predict the abundances of six major oxides (TiO2, FeO, Al2O3, MgO, CaO, SiO2) from eight-band SELENE (KAGUYA) Multiband Imager reflectance, using laboratory-measured abundances of 115 lunar soil samples from 55 Apollo, Luna and Chang'e-5 sampling sites as ground truth. The model was applied across the Multiband Imager data to produce global oxide abundance maps, and a second model trained without the Chang'e-5 samples was produced for comparison. Against measured sample abundances the reported RMSEs of this work were smaller than those of the Clementine UVVIS, Lunar Prospector GRS and Diviner CF products, while the FeO RMSE was 0.0027 higher than that of Chang'e-1 IIM. The maps were then used to derive an Mg# map, to outline the lunar maria, Feldspathic Highland and South Pole-Aitken units, and to define 26 young mare basalt compositional units.
How AI was used
Multiband Imager reflectance data at 59 m/pixel were shade-corrected and reduced to eight wavebands, with two spectral angle parameters computed to suppress optical maturity effects, and were paired with measured oxide abundances at the Apollo, Luna and Chang'e-5 sampling sites. A 1D convolutional neural network with five convolution blocks (channels 64, 64, 128, 128, 256; 1x3 kernels, stride 2 in the second and fourth blocks) feeding one fully connected layer was trained in supervised fashion to regress the six oxide abundances from each pixel's spectral sequence, using the Adam optimizer, weight decay, a capped epoch count, full-batch updates and early stopping. Training followed a two-stage scheme: leave-one-out cross-validation produced one best model per fold, and the weights of those models were averaged to form the final predictor, which was then run over the imagery to generate the oxide maps. Ablation over model size, learning rate and weight decay, and comparison with Extreme Learning Machine and Generalized Operational Perceptrons models trained under the same LOOCV configuration, were used to settle the configuration. Two training sets were used throughout, one including and one excluding the Chang'e-5 sample abundances, with identical network parameters.
The shape of the work
Structural · the record, drawn
no AI
Compile returned-sample ground truth
Obtaining raw data, whether by measurement, download or retrieval.
a total of 115 lunar soil samples acquired from 55 lunar sampling sites of Apollo, Luna, and Chang’e-5 landing regionswhere the paper describes this · verbatim
no AI
Pre-process SELENE MI spectra
Cleaning, filtering, normalising or labelling data already obtained.
the MI reflectance data have a resolution of 59 m/pixel after topography shadow correctionwhere the paper describes this · verbatim
AI
Train 1D CNN inversion model with leave-one-out folds and weight averaging
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
For each fold, one sample was selected as the validation set, and the remaining samples were adopted to train the model.where the paper describes this · verbatim
AI
Infer global oxide abundance maps
Running a trained model over new data to predict, classify or score. The AI stood in for statistical model.
The prediction module predicts the six major oxide abundances corresponding to the spectral observationwhere the paper describes this · verbatim
AI
Ablation and baseline model comparison under LOOCV
Testing outputs against ground truth. Its result feeds back into an earlier step.
we compared it with the standard Extreme Learning Machine (ELM) and a new MLPs variant, namely Generalized Operational Perceptrons (GOPs)where the paper describes this · verbatim
no AI
Compare maps with in situ data and published chemistry maps
Testing outputs against ground truth.
Two no-sample returned lunar landing site regions, i.e., Chang’e-3 and Chang’e-4, are selected to demonstrate the validity of the Chang’e-5 inversion resultswhere the paper describes this · verbatim
no AI
Partition geologic and young mare basalt compositional units
Extracting understanding from model behaviour.
Twenty-six young mare basalts compositional units (U1-U26) were defined and mapped in this study.where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's product is the set of global oxide abundance maps, every value of which is output by the trained 1D CNN inversion model; the geologic-unit conclusions rest on those maps.
a deep learning (DL) algorithm, i.e., a 1D convolutional neural network, was designed to establish an oxide inversion modelswhere the paper describes this · verbatim
the leave-one-out cross-validation (LOOCV) was adopted for evaluating the generalization ability of the inversion methodwhere the paper describes this · verbatim
a PC workstation (Intel(R) Xeon(R) Platinum 8352Y CPU @ 2.20GHz with 128 GB of RAM and NVIDIA GeForce RTX 3090 Graphics Processing Unit)where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of 1D convolutional neural network inversion model (this work)Which version of the model was used is not stated.
- Version of Extreme Learning Machine (ELM)Which version of the model was used is not stated.
- Version of Generalized Operational Perceptrons (GOPs)Which version of the model was used is not stated.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00209, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error