Department of Space / ISRO

Central Government


Chandrayaan 2: AI-powered ‘Pragyan’ Rover

On 22 July 2019, ISRO launched Chandrayaan 2 spacecraft into an earth orbit as part of the second lunar mission. Geosynchronous Satellite Launch Vehicle Mark-III (GSLV Mk -III) is a three-stage launch vehicle, which carried the spacecraft, comprised of the Orbiter, the Lander Vikram and the Rover Pragyan. The Pragyan Rover was placed strategically among the imperative areas of GSLV Mk-III and carried some key responsibilities/functionalities. The rover is a six-wheeled robotic vehicle and is capable of conduct in-situ payload experiments. It is powered by AI tools and frameworks, uses solar energy for its functioning and can communicate only with the Lander. The Pragyan Rover payloads consist of Alpha Particle X-ray Spectrometer (APXS) and Laser Induced Breakdown Spectroscope (LIBS).

₹987 crore

Learn more about the Initiative

Structural health monitoring through classification of strain patterns using artificial neural network

In November 2018, ISRO published a compendium for preparing project proposals by universities, which mentioned about programs and research areas in space such as launch vehicle, satellite communication, earth observations, space sciences, and meteorology. The Vikram Sarabhai Space Centre (VSSC) introduced structural health monitoring technology through classification of strain patterns using artificial neural network (ANN). The technology increases safety and reduces maintenance costs of high-performance composite structures used in aircraft and re-entry vehicles. ANN helps in detection of damages such as fibre failure, matrix cracking, de-laminations, skin-stiffener de-bonds in composite structures. It is used by ISRO to classify sensor malfunctioning and structural failures based on strain patterns of healthy and unhealthy structures. ISRO demanded an analytical study for the adopted methodology in its compendium.

Learn more about the Initiative

Respond Basket program by Capacity Building Programme Office

In November 2018, the Capacity Building Programme Office of ISRO introduced Respond Basket which comprised of approximately 150 research areas, out of which 13 are on AI and machine learning. The program is focused on key developments with AI and machine learning such as design and development of ANFIS controller for optimised control in Cryogenic High-Altitude Test facility of ISRO Propulsion Complex, algorithms for earth observation data processing, urban spatial growth modelling, mapping elevation using autonomous UAVs in swarm, object tracking and recognition, use of expert systems, genetic algorithms, NLP, robotics and fuzzy logics in mission operations, remote sensing based forestry and ecology, change prediction and modelling of dynamic natural ecosystem using image processing and remotely sensed data, target identification from MOTR radar data, online spatial data analysis and processing, software framework development, estimation and correction of spacecraft jitter to improve geometric performance of satellite images, and operations research/optimal control.

Learn more about the Initiative

Focus on geospatial technology-based services

In June 2018, Antrix Corporation (the commercial arm of ISRO) and SatSure Analytics (a satellite data analytics company which uses big data and machine learning technologies) signed an MoU to increase the penetration of geospatial technology-based services and to develop analytics products in various sectors such as agriculture, banking and financial services, social infrastructure, energy, and telecommunications. The partnership aims to develop synergies through technical advisory services, joint project execution, and national capacity building in the geospatial data analytical industry in India. It will also form a framework for effective collaboration with national space program and new home-grown startups with big data, machine learning, satellite imagery, and remote sensing capabilities.

Learn more about the Initiative

AI-enabled monitoring system for forest conservation

The National Remote Sensing Centre (NRSC), which ISRO has designed and developed, is a monitoring system to observe forest cover change and combat deforestation by leveraging optical remote sensing, geographic information system, AI, and automation technologies. The monitoring system allows experts to detect small-scale deforestation and improve the frequency of reporting. It also enables scientists to process satellite imagery faster and reduces the time frame for new reports from one year to one month. NSRC aims at preventing negative changes in the green cover and protection of wildlife. The NRSC technology makes it possible for monitoring forest cover changes over small areas of one hectare by improving the resolution from 50 meters to 30 meters through optical remote sensing which provides insights into the smallest of deforestation activity. 

Learn more about the Initiative

Gateway to Indian earth observation

ISRO launched an open data archive geo-portal called ‘Bhuvan,’ which provides visualization services and earth observation data to users in public domain, and offers remote sensing applications. The portal uses a crowdsourcing approach for collecting point-of-interest data and acts as a platform to host government data such as of the forest departments. It allows users to explore 2D and 3D representation of the surface of the earth, pest surveillance, disaster services, high-resolution imagery of cities, etc. States like Punjab, Karnataka, Himachal Pradesh and Andhra Pradesh use Bhuvan for specific applications in forestry, tourism, municipal, GIS, geo-tagging, etc.

The geo-portal is now enriched with 20 ministry portals and 30 state portals and assists government programs such as Integrated Watershed Development Program (Srishti-Drishti), National Mission for Clean Ganga, AMRUT, etc. The portal also enables latest tools and techniques such as data discovery, metadata display, database creation and maintenance, smart city programs, and surveillance and monitoring systems.

Learn more about the Initiative

Multi Object Tracking Radar (SDSC-SHAR)

The challenge of building Space object tracking solution to build successful sustenance of satellites through difficult terrain of open space with millions of unknown objects that could impact every ISRO sponsored mission.

The objective is to build Multi Object Tracking Radar. Detection of low RCS targets with high speed and complex motions has been receiving a growing attention and significant research efforts in the modern radar field especially in space debris tracking. Real-time implementation long term coherent integration by compensating phase fluctuation among different sampling pulses for high performance embedded computing platforms which involves searching the velocity fold factor, estimating acceleration, jerk values and performing CLEAN algorithm for integrating multiple targets has to be developed.

a. ISRO first developed Target identification using machine learning algorithms from MOTR radar data.

  • Radar data consists of Range, Azimuth, Elevation and Signal to Noise Ratio (SNR). From Range and SNR correlation target size can be classified. From SNR variation alone in a single-track duration, target nature can be established. Using Machine Learning algorithms, a model should be trained on radar tracked data (Range, Azimuth, Elevation and SNR). The trained model should identify a target nature (controlled or uncontrolled) and size. Using standard libraries in Python Machine Learning Algorithms have become realizable models

b. Followed by development of “Real time JPDA & MHT based data association in dense multi target tracking environment”.

  • MOTR has implemented Linear Kalman Filter (LKF) and Extended Kalman Filter (EKF) for tracking multiple targets simultaneously and Simple Nearest Neighborhood (SNN) based data association algorithm to associate target returns with the target being tracked. SNN data association algorithm gives a better result in tracking multiple targets when the targets being tracked are spatially separated. When multiple targets are very closer SNN algorithm gives poor result. It also fails in situation like targets cross over and co traveling of two targets. To overcome this situation Probability based Data Association (PDA) methods like Joint Probability Data Association (JPDA) and Multiple Hypothesis Tracking (MHT) algorithms are used. Since these algorithms uses probability-based algorithms these are complex incorporated to SNN. Hence these algorithms are mostly used in offline analysis.

c. Last critical element was to study, monitor and develop algorithm using historic data for “Space debris RCS estimation and dynamics characterisation from MOTR Space debris”

d. ISRO studied and designed algorithm based on received signal from the target gives us the information of the target like its dynamics spin, its size and RCS. These characteristics of the debris need to be catalogued, to compute its drag coefficient, and its life time assessment

Timeline: 12 to 24 months 

Image Processing and Pattern Recognition (IIRS)

In 1980s and 1990s ISRO’s the challenge was to build efficient and cost neutral image processing and pattern recognition solution for upcoming missions for next decade. Hence, Unmanned Image Processing and Pattern Recognition (IIRS).

a. ISRO leveraged Artificial Neural networks (ANN) which is generic name for a large class of machine learning algorithms, most of them are trained with an algorithm called back propagation. ISRO’s team used various path to explore various deep leaning algorithms in various applications of earth observation data like; self-learning based classification, prediction, multi-sensor temporal data in crop/forest species identification, remote sensing time series data analysis.

b. ISRO followed below approach and module-based solutions to design and develop state of art IIRS solution for future:

  • Super resolution approaches for hyperspectral remote sensing images (IIRS) – To enhance spatial resolution of the hyperspectral data, it needs to be fused with other sensor data of high spatial resolution. This results in enhanced spatial information but with loss of information. To obtain enhanced spatial information in hyperspectral data without using external data super resolution technique can be applied. The technique could be single frame super resolution technique, which converts hyperspectral data from low resolution to high resolution and can be utilized for detailed land cover analysis.
  • Subpixel target detection using spectral unmixing of hyperspectral data (IIRS) – Hyperspectral sensors are able to provide unprecedented spectral and radiometric excellence in the data sets. These datasets are very important for identifying features and discriminating various materials. Issues related to hyperspectral data processing include endmember extraction and subpixel analysis, which are essential for land cover information extraction. There are various automatic and semiautomatic techniques developed to extract endmembers and some of them relies on the existence of relatively pure pixels. Purest endmember can be extracted from the data collected by airborne hyperspectral sensors. Spectral unmixing in each pixel of hyperspectral data deals with the fact of decomposition of spectra present into its corresponding constituent components. Linear and nonlinear techniques of spectral unmixing allows proper subpixel analysis of the earth surface utilizing hyperspectral data.
  • Enhanced land cover information extraction from high resolution hyperspectral data using object-based technique (IIRS) – Satellite image classification for land-cover information extraction using high spectral and spatial resolution data is challenging task for traditional pixel-based classification approaches. The pixel-based classification approach only utilizes spectral information of the pixels to classify the image. Normally different physical objects have different spectral information and it is easy to differentiate using pixel-based classification. The ability of the approach is limited when objects have similar spectral information. Under this circumstance the image are not classified correctly. To separate the objects with similar spectral information object-based image analysis can be used which uses spatial as well as textural information. This approach segments the pixels into objects according to the homogeneity of the image and classify the image by treating objects as a whole
  • Advanced sensor models for optical & microwave data Geo-referencing (NRSC) – Georeferencing is the process of assigning spatial location to each pixel of an image using sensor models or GCPs. Precise georeferencing is a major issue especially in high resolution optical and microwave imagery. Presently, rigorous sensor models and rational function models are widely in used in optical imagery geo referencing. For microwave sensors the range Doppler method is generally employed for assigning precise geo location. This method corrects the imagery for foreshortening and layover effects by utilizing the topology, orbit and velocity measurements from satellite and assigns a geolocation to each pixel. We welcome proposals that defines new methods/models and implementation of any of the above algorithms both for optical and microwave imagery.
  • Atmospheric correction procedures implementation for Visible & NIR & HySI (NRSC) – Atmospheric correction is key image processing step to retrieve surface reflectance values from spectra recorded by remote sensing space borne sensors. This further helps in standardizing physical variables, thus facilitating comparisons across time series of such variables. Presently, Atmospheric correction of all bands of Resourcesat-2 AWIFS and LISS III sensors is being carried out utilizing water vapor, ozone data products from MODIS, Aerosol optical depth from INSAT satellites using 6S RTF algorithm. The atmospheric correction has improved our estimation of normalized difference vegetation index by a factor of 50% with respect to TOA. There are possibilities of improving these estimations further by modelling the Bidirectorial reflectance distribution functions, using reflectance references from Drones or physics-based models. We invite proposals in any of these areas that aim to minimize the influence of atmospheric effects in estimation of physical variables.

Timeline: 24 to 36 months 

Autonomously Navigating Robot for Space Mission (IISU)

ISRO’s challenge was to build and send unmanned robots to help fetch critical space information in multiple missions throughout the year.

A half Vyomnoid with Sensing and perception of surroundings with 3D vision and Dexterous manipulative abilities to carry out defined crew functions in an unmanned mission or assist crew in manned missions. Design & Realization of FULL Vyomnoid with features that include full autonomy with 3D vision, dynamically controlled movement in zero ‘g’, Artificial Intelligence / Machine Learning enabled real time decision making with vision optimization and path planning algorithms.

ISRO leveraged state of technologies to design and develop following:

  • Sensing & Perception
  • Dexterous Manipulation
  • Hierarchical Control System
  • Artificial Intelligence enabled Path Navigation algorithms

Timeline: 10 to 12 months