Move over, NASA, this Jharkhand AI breakthrough could guide future space missions

At BIT Mesra in Ranchi, a three-woman team has trained AI to detect and analyse lunar craters. The ISRO-backed work could support crater dating, navigation planning and future Moon landing missions.

In Short

• An ISRO-funded project began after academia was asked to automate crater counting
• Researchers trained models on Chandrayaan and NASA datasets over three years
• The system catalogued impacts as small as 200 metres accurately

Hello, can you hear me? The audio breaks into digital static and the video freezes, a frustrating reminder of the technological friction that exists in long-distance conversations to this day. But for the three women sitting on the other end of the line at the Birla Institute of Technology (BIT) Mesra in Ranchi, a spotty Wi-Fi signal is only a minor hurdle. For three years, their eyes have been fixed on the Moon.

“Our main mission was deep learning lunar crater detection,” Dr Sanchita Paul, an associate professor in the Department of Computer Science, tells India Today Tech. Beside her sit Dr Mili Ghosh from the Remote Sensing Department and Senior Research Fellow, Mimansa Sinha. Together, this three-woman team has spent three years doing what was once a job only experienced geologists could do. They have been teaching artificial intelligence how to look at the Moon and map its topography, simultaneously rewriting the playbook for future lunar landings.

The project, funded by the Indian Space Research Organisation (ISRO), began in February 2023, though its seeds were planted during a Chandrayaan-2 data analysis workshop a year prior. ISRO had a problem. Historically, counting and analysing lunar craters was done manually. Geologists would sit for long hours, tracing the rims of craters on satellite images. It was exhausting and painfully slow. ISRO issued a challenge to academia: We need automation.

For Dr Mili Ghosh, who already had experience with the Chandrayaan-1 mission, the opportunity was a natural fit. But for Dr Sanchita Paul, a computer scientist accustomed to standard digital images, stepping into the cosmic arena was a paradigm shift.

“This is the first time computer science has collaborated with ISRO at our institute,” Dr Paul reflects. “We computer science people generally work on standard PNG or JPEG images. But satellite data uses Digital Elevation Models (DEM). It is a completely different, highly complex TIFF dataset. It was a massive learning curve figuring out how these 3D images could be passed through machine learning and deep learning techniques to finally identify a crater.”

All in on AI

Teaching an AI to recognise a crater sounds simple until you factor in the Moon’s massively uneven surface. The team worked with a dizzying array of advanced neural networks on datasets pulled from NASA’s Lunar Reconnaissance Orbiter (LROC) and India’s Chandrayaan missions.

They tested everything: U-Net architectures, various iterations of the YOLO (You Only Look Once) real-time object detection models, and multiple convolutional neural networks. It was a relentless process of trial and error.

“AI is basically a prediction system,” Dr Paul explains. “We cannot say we are 100 per cent accurate. If somebody says an AI system automatically detects something with absolute perfection, that is incorrect. It’s a trial-and-error method. We try different techniques, observe the results, and select the best.”

By fusing existing frameworks and hyper-tuning them to fit their specific lunar dataset, the team eventually landed on an architecture combining Mask R-CNN with a ResNet-101 backbone.

“Brightness, resolution, everything depends on the technique you are using and how strong your network backbone is,” says Dr Paul. “Our backbone is ResNet-101, a 101-layer deep architecture. It is an incredibly extensive, powerful tool for feature extraction.”

The BIT Mesra team didn’t just want to find craters, it wanted to understand them. Once the Mask R-CNN model isolates a crater, the data is pushed into a customised software extension the team developed from scratch. It is called the CraterMorpho. Built as an automated toolbox for ArcGIS Pro, this homegrown software instantly extracts the crater’s morphometry including its precise diametre, depth, slope inclination, circularity, and the topographic roughness of its walls.

The AI proved to be remarkably efficient. It successfully mapped “sub-kilometre” craters, tiny impacts less than 0.5 kilometres wide that standard automation routines routinely miss. It mastered the chaotic geometry of craters inside other craters and overlapping impacts. To test the limits of their tool, the researchers narrowed their focus to a specific, notoriously complex study area on the Moon’s Aristarchus Plateau: Vallis Schrteri or Schroter's Valley.

“Vallis Schrteri is a massive, sinuous volcanic rille,” Dr Paul explains. “Above it, a spectacular large crater is visible, which is why astronomers call it the ‘Cobra Head.’ We took a dataset of 35 high-resolution images of this area and generated its crater catalogue. Why? Because it has historically experienced the most intense impacts from meteoroids, asteroids, and ancient volcanic eruptions.”

The resulting catalogue is a goldmine for planetary science. It identifies and measures craters down to a mere 200 metres in size, charting their exact coordinates of latitude, longitude, and altitude.

Decoding lunar time

For the average young professional or tech enthusiast, the immediate question might be: Why do we care so much about potholes on the Moon? The answer, as it turns out, is that craters are the universe’s oldest grandfather clocks.

“The Moon has no atmosphere,” Mimansa Sinha explains. “Because there are no weathering agents like liquid water or wind, everything that hits the surface remains pristine. It is preserved over geological timescales. By automating the crater-counting process, we can calculate the absolute age of different parts of the lunar surface.”

Furthermore, this tool can help in change detection. By feeding the AI historical lunar data from 50 years ago and comparing it to fresh imagery captured by Chandrayaan-2, scientists can instantly see if the crater count has increased. If the numbers change, it means that a specific region of the Moon is still actively experiencing impact cratering today.

Understanding the structural integrity of these craters also holds the key to the future of human spaceflight. CraterMorpho automatically classifies craters into three categories: fresh, moderately degraded, and highly degraded.

“Just like we define the age of wood based on its rings, the more eroded and degraded a crater is, the older it is,” says Dr Paul.

By identifying where the smooth, ancient, degraded terrain lies and where the sharp, treacherous, freshly formed craters are hidden, the AI provides a critical survival map for future rovers and crewed landing missions. It can be the ultimate tool for autonomous navigation planning and landing site analysis.

The new vanguard of Indian research

The brilliance of this project hasn't gone unnoticed by the global scientific community. The team’s findings have been peer-reviewed and published in the world's leading planetary science journals, including Icarus, Scientific Reports, and Planetary and Space Science.

For an institute nestled in Ranchi, Jharkhand, competing on the global stage of space exploration is a testament to a changing academic culture in India. The growing push by institutions like BIT Mesra to balance rigorous classroom teaching with cutting-edge, funded research is bearing fruit.

“Our honourable Vice Chancellor has made it clear that we must excel in research,” notes Dr Mili Ghosh, pointing to the supportive ecosystem cultivated by the university’s Dean of Research office. It’s an environment that allows scientists to dream big and move fast. Case in point: Mimansa Sinha fulfilled all the stringent criteria for her PhD submission in a breakneck two and a half years, a timeline almost unheard of in deep-tech research.

What makes the team truly formidable, however, is their sheer versatility. Dr Paul isn't just looking at the Moon. She is simultaneously serving as the Principal Investigator on a project analysing human brainwaves (EEG/ECG signals) to assist wheelchair control, and another dedicated to using AI to predict the spread of vector-borne diseases on Earth.

“There is no end goal,” she says simply when asked about juggling such wildly diverse, high-stakes domains. “As a researcher, we never have an end goal. We always want to find out something new, do something new, and try to achieve something further.”

Looking to the dark side

As the global space race heats up, with nations vying to establish a permanent presence on the lunar surface, the work coming out of Ranchi is far from over. The team has already submitted proposals for the Chandrayaan-3 dataset, eager to put their optimised Mask R-CNN models on the shadow-draped terrain of the lunar South Pole.

Further on the horizon lies Chandrayaan-4, a highly anticipated sample-return mission designed to bring physical pieces of the Moon back to Earth. For Dr Ghosh and the team, that mission represents the ultimate scientific pay-off.

“In remote sensing, you cannot predict everything purely through satellite image analysis,” Dr Ghosh says, looking forward. “You need physical validation. The sample-return mission will provide that validation. We are incredibly hopeful to be working alongside ISRO when that happens.”

The journey from the classrooms of Ranchi to the craters of the Moon is a powerful story of how the democratisation of AI is reshaping our world and even the world(s) beyond. Armed with Python code, open-source geospatial tools, and an unyielding curiosity, three women have built a digital lens that brings the cosmos into sharper focus. And as humanity prepares to take the next leap into the great beyond, the maps guiding their feet may very well be drawn by intelligence trained in the heart of Jharkhand. Whatever is the case, things are bound to only go up, up and away from here.

- Ends