💻 computer science

Towards Seamless Lunar Mosaics: Deep Radiometric Normalization for Cross-Sensor Orbital Imagery Using Chandrayaan-2 TMC Data

This paper presents a deep learning-based radiometric normalization framework using a conditional generative adversarial network (cGAN) to effectively eliminate inconsistencies and generate seamless, high-fidelity lunar mosaics from heterogeneous multi-mission orbital imagery, primarily Chandrayaan-2 TMC data.

Original authors: Pratincha Singh, Jai Gopal Singla, Prashant Hemrajani, Nitant Dube, Amithabh, Hinal Patel

Published 2026-04-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pratincha Singh, Jai Gopal Singla, Prashant Hemrajani, Nitant Dube, Amithabh, Hinal Patel

Original paper licensed under CC BY 4.0 (http://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 trying to create a giant, single picture of the Moon by gluing together thousands of smaller photos taken by different spacecraft over many years. This is what scientists call a "lunar mosaic."

However, there's a big problem: these photos don't look like they belong together. Some are bright, some are dark, some have a blue tint, and others look washed out. It's like trying to make a collage using photos taken at noon, photos taken at sunset, and photos taken with three different types of cameras. The result is a patchwork quilt with obvious, ugly seams where the colors don't match.

This paper presents a new, smart way to fix those seams using Artificial Intelligence (AI). Here is how they did it, explained simply:

The Problem: The "Patchwork Quilt" Effect

The researchers used data from India's Chandrayaan-2 mission (a high-resolution camera called TMC) and Japan's SELENE mission. When they tried to stitch these images together, the lighting looked wrong.

Old Way: Previously, scientists used simple math tricks (like "histogram matching") to force the colors to look similar. Think of this as trying to fix a mismatched quilt by just squinting your eyes or using a filter on your phone. It helps a little, but the seams are still visible, and the details get blurry.

The Solution: The "AI Art Teacher"

The authors built a deep learning system (a type of advanced AI) to act like a master art teacher. They didn't just tell the computer to "make it brighter"; they taught it to understand how the Moon actually looks under consistent lighting.

The Reference Photo: They used a "gold standard" map of the Moon created by NASA's LROC WAC camera. This map is perfectly consistent. Think of this as the "perfect reference painting" that the AI is trying to copy.
The Student (The Generator): The AI uses a neural network called U-Net. Imagine this as a student artist who looks at a messy, mismatched photo of the Moon and tries to redraw it to look exactly like the perfect reference painting, while keeping all the craters and rocks in the right place.
The Critic (The Discriminator): The AI also has a "critic" (called a PatchGAN). This critic looks at small pieces of the new image and asks, "Does this look like a real, consistent part of the Moon, or does it look fake?" If the student makes a mistake (like a weird seam), the critic says, "Try again."
The Learning Process: The student and the critic play a game. The student tries to fool the critic, and the critic tries to catch the student. Over time (about 150 rounds of training), the student gets so good at matching the reference that the critic can't tell the difference.

How They Handled the Whole Moon

The Moon is huge, and the computer can't look at the whole thing at once (it would run out of memory).

The Puzzle Piece Strategy: The researchers cut the giant Moon map into smaller square "puzzle pieces" (patches).
The Overlap Trick: They made these puzzle pieces overlap slightly. When the AI fixes one piece, it blends the edges with the neighbors. This ensures that when you put the pieces back together, there are no visible cracks or lines.

The Results: A Seamless View

After the AI finished its work, the results were impressive:

Smoother Transitions: The ugly lines where different photos met disappeared. The brightness flowed naturally from one area to the next.
Better Details: Unlike the old methods, which sometimes blurred the surface, this AI kept the sharp details of craters and ridges intact.
Proof: The researchers measured the quality using standard math scores (PSNR and SSIM). The AI's version scored much higher than the old methods, proving it was much closer to the "perfect" reference map.

What They Didn't Do

It is important to note what this paper didn't do:

They didn't fix the geometry (if two photos are slightly shifted or tilted, this AI doesn't fix that; it only fixes the colors and brightness).
They didn't test this on other planets yet; they only tested it on the Moon.
They didn't claim this works instantly for any planet; it specifically needed a "perfect reference map" (like the NASA one) to learn from.

In a Nutshell

This paper shows that by using a smart AI "teacher" and "critic" to learn from a perfect reference map, we can take a messy collection of Moon photos from different missions and stitch them together into one seamless, high-quality map without the ugly seams. It's like using a super-smart editor to make a collage look like it was taken in a single, perfect moment.

1. Problem Statement

The construction of seamless, large-scale lunar mosaics from multi-mission orbital imagery is hindered by significant radiometric inconsistencies. These discrepancies arise from:

Variable Illumination Geometry: Differences in solar angles during image acquisition.
Sensor Characteristics: Variations in response between different missions (e.g., ISRO's Chandrayaan-2 vs. JAXA's SELENE/Kaguya).
Acquisition Conditions: Differences in exposure and imaging geometry.

These factors result in visible seam artifacts, tonal discontinuities, and brightness mismatches in final mosaics. Traditional methods, such as global histogram matching and block-based tone balancing, often fail to handle the complex, nonlinear variations found in heterogeneous multi-source datasets, leading to poor visual quality and reduced scientific interpretability.

2. Methodology

The authors propose a Deep Learning-based Radiometric Normalization Framework utilizing a Conditional Generative Adversarial Network (cGAN). The pipeline consists of three main stages:

A. Data Preparation & Preprocessing

Datasets: The primary input is Chandrayaan-2 Terrain Mapping Camera (TMC) data (5m resolution), supplemented by SELENE (Kaguya) imagery (8m resolution) to fill coverage gaps.
Reference: LROC Wide Angle Camera (WAC) data (100m resolution) serves as the photometrically consistent ground truth/reference.
Preprocessing: Raw images undergo Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance local contrast.
Patch Extraction: Large mosaics are divided into overlapping patches (512×512 or 1024×1024 pixels) to facilitate efficient training and preserve local structural continuity.

B. Model Architecture (cGAN)

The core of the framework is a cGAN designed to learn a nonlinear mapping from the input mosaic to the reference distribution:

Generator: Based on a U-Net architecture. It uses an encoder-decoder structure with skip connections to extract multi-scale features and preserve spatial details (e.g., crater rims, ridges) while normalizing radiometry.
Discriminator: A PatchGAN discriminator. Instead of evaluating the whole image, it classifies local patches as real (reference) or fake (generated), enforcing high-frequency consistency and reducing local seam artifacts.

C. Training Strategy & Loss Functions

Objective Function: The model minimizes a combined loss function:
- Adversarial Loss ( $L_{cGAN}$ ): Ensures the generated output is indistinguishable from the reference distribution.
- L1 Loss ( $L_{L1}$ ): Enforces pixel-level structural similarity between the generated image and the reference.
- Total Loss: $L_{total} = L_{cGAN} + \lambda L_{L1}$ (where $\lambda=100$ ).
Inference: A sliding window approach with overlap-aware inference is used. Overlapping patches are reconstructed using weighted averaging (Gaussian blending) to ensure seamless transitions across tile boundaries.

3. Key Contributions

Deep Learning Framework for Multi-Mission Mosaics: Introduces a novel cGAN-based approach specifically tailored for normalizing heterogeneous lunar datasets (Chandrayaan-2 TMC + SELENE) against a reference (LROC WAC).
Nonlinear Radiometric Mapping: Moves beyond linear histogram matching by learning complex, nonlinear transformations that account for varying illumination and sensor responses.
Scalable Patch-Based Strategy: Implements an overlap-aware inference mechanism that allows the processing of large-area mosaics while maintaining structural continuity and minimizing boundary artifacts.
Comprehensive Evaluation: Provides both quantitative (PSNR, SSIM) and qualitative (visual, histogram) evidence of the method's superiority over traditional techniques.

4. Experimental Results

The framework was trained for 150 epochs on an NVIDIA H100 GPU.

Quantitative Performance:
- The model achieved peak performance at Epoch 125, with a PSNR of 30.55 dB and an SSIM of 0.9876.
- These metrics indicate a high degree of radiometric similarity and structural preservation compared to the LROC WAC reference.
- The results consistently outperformed traditional histogram-based normalization techniques.
Qualitative Observations:
- Visual Consistency: The deep learning-generated mosaics showed significantly reduced brightness discontinuities and smoother transitions between tiles compared to the baseline.
- Histogram Alignment: The intensity distribution of the generated mosaics closely matched the reference WAC mosaic, whereas traditional methods exhibited irregular peaks and broader distributions.
- Feature Preservation: Fine geological structures (crater rims, ejecta patterns) were preserved, demonstrating that the normalization did not blur critical surface details.
Training Stability: The adversarial training showed stable convergence, with the discriminator loss stabilizing around 0.4, indicating balanced training without mode collapse.

5. Significance and Future Work

Significance:

Scientific Utility: High-fidelity, seamless mosaics are critical for accurate geomorphological analysis, surface change detection, and future mission planning.
Scalability: The proposed method offers a scalable, automated solution for generating planetary maps from heterogeneous data, reducing the need for extensive manual tuning required by conventional methods.
Generalizability: While focused on the Moon, the framework demonstrates the potential of deep generative models for planetary image processing across different celestial bodies.

Limitations & Future Directions:

Reference Dependency: The method relies on the availability of a high-quality radiometric reference dataset (LROC WAC), which may not exist for all planetary bodies.
Patch Constraints: Fixed-size patch training may limit the capture of very large-scale illumination gradients extending across multiple tiles.
Geometric Alignment: The current framework addresses radiometry only; future work aims to integrate geometric correction and parallax removal.
Resolution: Future research will target full-resolution Chandrayaan-2 TMC (5m) and Orbiter High Resolution Camera (OHRC) data, potentially incorporating physics-informed learning strategies.

In conclusion, this paper demonstrates that deep generative models, specifically cGANs with U-Net generators and PatchGAN discriminators, are highly effective at suppressing seam artifacts and harmonizing radiometry in multi-mission lunar mosaics, paving the way for next-generation planetary mapping.