Muthukumaran Ramasubramanian
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Muthukumaran Ramasubramanian has a total of 105 co-authors. Here are the top co-authors by number of shared publications:
- Rahul Ramachandran
NASA Marshall Space Flight Center, Huntsville, AL, USA
8 shared publications - Iksha Gurung
University of Alabama in Huntsville
7 shared publications - Manil Maskey
NASA Marshall Space Flight Center
6 shared publications - Brian Freitag
National Aeronautical and Space Administration Marshall Space Flight Center Huntsville AL USA
4 shared publications Aaron Kaulfus
NASA Marshall Space Flight Center Huntsville AL USA
2 shared publications- Gabriele Cavallaro
University of Iceland
2 shared publications
Latest Publications
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 11762-11773
INDUS-SDE: A Language Model for Scientific Content Curation and Discovery
Scientific content curation from organizing, tagging, classifying, to connecting scientific knowledge is essential for accelerating discovery but cannot scale through manual expert effort alone. We present INDUS-SDE, a language model designed to learn from expert-curated scientific content and extend curation capabilities at scale. Our key innovation, Weighted Dynamic Masking, prioritizes scientific terminology during pretraining by combining random masking with targeted masking of domain-relevant keywords. This approach addresses the sparsity problem inherent in web-sourced scientific corpora, where domain-relevant terminology is sparse relative to structural noise such as navigation elements, headers, and boilerplate text, resulting in low signal-to-noise ratios that challenge standard pretraining. By prioritizing scientific keywords, INDUS-SDE learns robust representations despite this content heterogeneity. We validate INDUS-SDE on NASA's Science Discovery Engine (SDE), a real-world platform aggregating scientific content across five NASA science divisions. INDUS-SDE substantially outperforms baseline models on masked language modeling and demonstrates strong performance on downstream curation tasks including metadata keyword recommendation and domain classification. INDUS-SDE matches the performance of some domain-specific models trained on significantly more in-domain data; this shows effective generalization from general scientific pretraining. We also introduce INDUS-SDE-ST, a sentence transformer that enables semantic scientific discovery. INDUS-SDE is currently deployed in NASA's SDE pipeline, accelerating curation workflows that previously required extensive manual effort. Our approach demonstrates that language models can effectively learn from expert curation workflows to scale scientific content organization.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 11762-11773
Open AccessINDUS-SDE: A Language Model for Scientific Content Curation and Discovery
Scientific content curation from organizing, tagging, classifying, to connecting scientific knowledge is essential for accelerating discovery but cannot scale through manual expert effort alone. We present INDUS-SDE, a language model designed to learn from expert-curated scientific content and extend curation capabilities at scale. Our key innovation, Weighted Dynamic Masking, prioritizes scientific terminology during pretraining by combining random masking with targeted masking of domain-relevant keywords. This approach addresses the sparsity problem inherent in web-sourced scientific corpora, where domain-relevant terminology is sparse relative to structural noise such as navigation elements, headers, and boilerplate text, resulting in low signal-to-noise ratios that challenge standard pretraining. By prioritizing scientific keywords, INDUS-SDE learns robust representations despite this content heterogeneity. We validate INDUS-SDE on NASA's Science Discovery Engine (SDE), a real-world platform aggregating scientific content across five NASA science divisions. INDUS-SDE substantially outperforms baseline models on masked language modeling and demonstrates strong performance on downstream curation tasks including metadata keyword recommendation and domain classification. INDUS-SDE matches the performance of some domain-specific models trained on significantly more in-domain data; this shows effective generalization from general scientific pretraining. We also introduce INDUS-SDE-ST, a sentence transformer that enables semantic scientific discovery. INDUS-SDE is currently deployed in NASA's SDE pipeline, accelerating curation workflows that previously required extensive manual effort. Our approach demonstrates that language models can effectively learn from expert curation workflows to scale scientific content organization.
- 17 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 13 (4) , 457-462
Earth Action in Transition: Highlights From the 2025 ESA–NASA International Workshop on AI Foundation Models for EO [Space-Agencies]
Over 850 people joined the first International Workshop on AI Foundation Model (FM) for Earth Observation (EO) co-organized by ESA and NASA 5-7 May 2025. Hosted at ESRIN (ESA’s Earth Observation Center, Italy), the event welcomed around 300 people on site, and an additional 550 online, with the promise that FMs can revolutionize EO and Earth sciences. The workshop marked a pivotal moment in aligning the EO and FM communities, fostering a shared commitment to developing open and trustworthy tools that support science discovery, operational applications, and prescriptive analytics. EO data is massive, complex and high dimensional requiring specific yet scalable AI architectures. The workshop emphasized the need for training and architecture enabling interpretability, explainability, and physical consistency. Coordination should be strengthened to minimize redundant development and to better leverage collective expertise. The focus has shifted from prototyping to real-world deployment, with FMs needing further design for integration into digital twins, dashboards, and edge platforms. Transparent benchmarking and user-driven evaluation are key to guiding model development and decision-making. In addition, parameter-efficient adaptation, neural compression, and embedding-based workflows offer promising paths for scaling EO analytics. While FMs show promise, their effectiveness remains context-dependent. The community debated whether to pursue universal models, specialized solutions, or mixtures of experts. The workshop envisioned the future of agentic AI in EO, with multi-agent system powered by EO FMs and vision-language models, that can dynamically reason and act on EO data. This shift from static pipelines to adaptive, smarter systems could redefine the future of EO. This paper summarizes key discussions and concludes with thought-provoking remarks.
- 17 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 13 (4) , 457-462
Earth Action in Transition: Highlights From the 2025 ESA–NASA International Workshop on AI Foundation Models for EO [Space-Agencies]
Over 850 people joined the first International Workshop on AI Foundation Model (FM) for Earth Observation (EO) co-organized by ESA and NASA 5-7 May 2025. Hosted at ESRIN (ESA’s Earth Observation Center, Italy), the event welcomed around 300 people on site, and an additional 550 online, with the promise that FMs can revolutionize EO and Earth sciences. The workshop marked a pivotal moment in aligning the EO and FM communities, fostering a shared commitment to developing open and trustworthy tools that support science discovery, operational applications, and prescriptive analytics. EO data is massive, complex and high dimensional requiring specific yet scalable AI architectures. The workshop emphasized the need for training and architecture enabling interpretability, explainability, and physical consistency. Coordination should be strengthened to minimize redundant development and to better leverage collective expertise. The focus has shifted from prototyping to real-world deployment, with FMs needing further design for integration into digital twins, dashboards, and edge platforms. Transparent benchmarking and user-driven evaluation are key to guiding model development and decision-making. In addition, parameter-efficient adaptation, neural compression, and embedding-based workflows offer promising paths for scaling EO analytics. While FMs show promise, their effectiveness remains context-dependent. The community debated whether to pursue universal models, specialized solutions, or mixtures of experts. The workshop envisioned the future of agentic AI in EO, with multi-agent system powered by EO FMs and vision-language models, that can dynamically reason and act on EO data. This shift from static pipelines to adaptive, smarter systems could redefine the future of EO. This paper summarizes key discussions and concludes with thought-provoking remarks.
- Open Access
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 98-112
INDUS: Effective and Efficient Language Models for Scientific Applications
Bishwaranjan Bhattacharjee, Aashka Trivedi, Masayasu Muraoka, Muthukumaran Ramasubramanian, Takuma Udagawa, Iksha Gurung, Nishan Pantha, Rong Zhang, Bharath Dandala, Rahul Ramachandran, Manil Maskey, Kaylin Bugbee, Michael M. Little, Elizabeth Fancher, Irina Gerasimov, Armin Mehrabian, Lauren Sanders, Sylvain V. Costes, Sergi Blanco-Cuaresma, Kelly Lockhart, Thomas Allen, Felix Grezes, Megan Ansdell, Alberto Accomazzi, Yousef El-Kurdi, Davis Wertheimer, Birgit Pfitzmann, Cesar Berrospi Ramis, Michele Dolfi, Rafael Teixeira De Lima, Panagiotis Vagenas, S. Karthik Mukkavilli, Peter W. J. Staar, Sanaz Vahidinia, Ryan McGranaghan, Tsengdar J. Lee. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 2024.
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 98-112
Open AccessINDUS: Effective and Efficient Language Models for Scientific Applications
Bishwaranjan Bhattacharjee, Aashka Trivedi, Masayasu Muraoka, Muthukumaran Ramasubramanian, Takuma Udagawa, Iksha Gurung, Nishan Pantha, Rong Zhang, Bharath Dandala, Rahul Ramachandran, Manil Maskey, Kaylin Bugbee, Michael M. Little, Elizabeth Fancher, Irina Gerasimov, Armin Mehrabian, Lauren Sanders, Sylvain V. Costes, Sergi Blanco-Cuaresma, Kelly Lockhart, Thomas Allen, Felix Grezes, Megan Ansdell, Alberto Accomazzi, Yousef El-Kurdi, Davis Wertheimer, Birgit Pfitzmann, Cesar Berrospi Ramis, Michele Dolfi, Rafael Teixeira De Lima, Panagiotis Vagenas, S. Karthik Mukkavilli, Peter W. J. Staar, Sanaz Vahidinia, Ryan McGranaghan, Tsengdar J. Lee. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 2024.
- 25 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 11 (3) , 129-132
A Summer School Session on Mastering Geospatial Artificial Intelligence: From Data Production to Artificial Intelligence Foundation Model Development and Downstream Applications [Technical Committees]
Provides society information that may include news, reviews or technical notes that should be of interest to practitioners and researchers.
- 25 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 11 (3) , 129-132
A Summer School Session on Mastering Geospatial Artificial Intelligence: From Data Production to Artificial Intelligence Foundation Model Development and Downstream Applications [Technical Committees]
Provides society information that may include news, reviews or technical notes that should be of interest to practitioners and researchers.
- Open Access
- 28 February 2023
- Published by American Geophysical Union (AGU) in Earth and Space Science
- Vol. 10 (3)
Tropical Cyclone Wind Speed Estimation: A Large Scale Training Data Set and Community Benchmarking
Tropical cyclones (TCs) cause significant disruptions to infrastructure and livelihood. The scale of loss due to TCs may be mitigated by prompt and accurate advisories about TC wind speed. Current advisories are consensus based and have a time delay of about 6 hr between each new update. As part of efforts to increase the frequency of wind speed estimations without compromising on the accuracy, we curate a high‐quality hurricane image and wind speed data set and develop a machine learning (ML) model for frequent wind speed estimation. The image and wind speed pairings are consolidated from different sources. The wind speeds are interpolated hourly to curate the TC wind speed estimation data set. The data set is supplemented with metadata that eases adoption of the data set by the ML community. We have also designed a competition to garner interest and encourage community involvement to build state‐of‐the‐art ML models to estimate wind speeds from satellite images of TCs. The competition winners were able to improve the wind speed estimation by almost 50% over the benchmark model.
- 28 February 2023
- Published by American Geophysical Union (AGU) in Earth and Space Science
- Vol. 10 (3)
Open AccessTropical Cyclone Wind Speed Estimation: A Large Scale Training Data Set and Community Benchmarking
Tropical cyclones (TCs) cause significant disruptions to infrastructure and livelihood. The scale of loss due to TCs may be mitigated by prompt and accurate advisories about TC wind speed. Current advisories are consensus based and have a time delay of about 6 hr between each new update. As part of efforts to increase the frequency of wind speed estimations without compromising on the accuracy, we curate a high‐quality hurricane image and wind speed data set and develop a machine learning (ML) model for frequent wind speed estimation. The image and wind speed pairings are consolidated from different sources. The wind speeds are interpolated hourly to curate the TC wind speed estimation data set. The data set is supplemented with metadata that eases adoption of the data set by the ML community. We have also designed a competition to garner interest and encourage community involvement to build state‐of‐the‐art ML models to estimate wind speeds from satellite images of TCs. The competition winners were able to improve the wind speed estimation by almost 50% over the benchmark model.
- 1 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 10 (4) , 273-277
Report on the IEEE GRSS Workshop on Remote Sensing Data Management Technologies in Geoscience 2022 [Technical Committees]
In recognition of emerging new data management technologies, the IEEE Earth Science Informatics (ESI) Technical Committee (TC) recently formed a new Working Group on Databases in Remote Sensing (DBRS). This is a report about the first workshop organized by the DBRS-WG to gather information about technologies that could effectively store, query, search, and analyze remote sensing data. This workshop called for research talks relevant to geoscience and remote data sensing technologies. The hybrid event, hosted at the Earth System Science Center (ESSC), University of Alabama in Huntsville (UAH), drew 16 submissions from around the world. The workshop also featured five invited speakers to talk about the advanced data management technologies and important application drivers. The WG believes that this inaugural event is a good start of a community around remote sensing data management.
- 1 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 10 (4) , 273-277
Report on the IEEE GRSS Workshop on Remote Sensing Data Management Technologies in Geoscience 2022 [Technical Committees]
In recognition of emerging new data management technologies, the IEEE Earth Science Informatics (ESI) Technical Committee (TC) recently formed a new Working Group on Databases in Remote Sensing (DBRS). This is a report about the first workshop organized by the DBRS-WG to gather information about technologies that could effectively store, query, search, and analyze remote sensing data. This workshop called for research talks relevant to geoscience and remote data sensing technologies. The hybrid event, hosted at the Earth System Science Center (ESSC), University of Alabama in Huntsville (UAH), drew 16 submissions from around the world. The workshop also featured five invited speakers to talk about the advanced data management technologies and important application drivers. The WG believes that this inaugural event is a good start of a community around remote sensing data management.
- Open Access
- 13 April 2022
- Published by Wiley in ESS Open Archive
Curating flood extent data and leveraging citizen science for benchmarking machine learning solutions
We present a labeled machine learning (ML) training dataset derived from Sentinel 1 C-band synthetic aperture radar (SAR) data for flood events. In this paper, we detail the steps to collect, pre-process, label, curate, and catalog the training dataset. Development of benchmark ML models and usage of the training datasets for a data science competition are also presented.
- 13 April 2022
- Published by Wiley in ESS Open Archive
Open AccessCurating flood extent data and leveraging citizen science for benchmarking machine learning solutions
We present a labeled machine learning (ML) training dataset derived from Sentinel 1 C-band synthetic aperture radar (SAR) data for flood events. In this paper, we detail the steps to collect, pre-process, label, curate, and catalog the training dataset. Development of benchmark ML models and usage of the training datasets for a data science competition are also presented.
- Open Access
- 27 March 2022
- Published by Copernicus GmbH
Language model for Earth science for semantic search
Recent advances in technology have transformed the Natural Language Technology (NLT) landscape, specifically, the use of transformers to build language models such as BERT and GPT3. Furthermore, it has been shown that the quality and the domain-specificity of input corpus to language models can improve downstream application results. However, Earth science research has minimal efforts focused on building and using a domain-specific language model. We utilize a transfer learning solution that uses an existing language model trained for general science (SciBERT) and fine-tune it using abstracts and full text extracted from various Earth science journals to create BERT-E (BERT for Earth Science). The training process utilized the input of 270k+ Earth science articles with almost 6 million paragraphs. We used Masked Language Modeling (MLM) to train the transformer model. MLM works by masking random words in the paragraph and optimizing the model for predicting the right masked word. BERT-E was evaluated by performing a downstream keyword classification task, and the performance was compared against classification results using the original SciBERT Language Model. The SciBERT-based model attained an accuracy of 89.99, whereas the BERT-E-based model attained an accuracy of 92.18, showing an improvement in overall performance.We investigate employing language models to provide new semantic search capabilities for unstructured text such as papers. This search capability requires utilizing a knowledge graph generated from Earth science corpora with a language model and convolutions to surface latent and related sentences for a natural language query. The sentences in the papers are modeled in the graph as nodes, and these nodes are connected through entities. The language model is used to give sentences a numeric representation. Graph convolutions are then applied to sentence embeddings to obtain a vector representation of the sentence along with combined representation of the surrounding graph structure. This approach utilizes both the power of adjacency inherently encoded in graph structures and latent knowledge captured in the language model. Our initial proof of concept prototype used SIMCSE training algorithm (and the tinyBERT architecture) as the embedding model. This framework has demonstrated an improved ability to surface relevant, latent information based on the input query. We plan to show new results using the domain-specific BERT-E model.
- 27 March 2022
- Published by Copernicus GmbH
Open AccessLanguage model for Earth science for semantic search
Recent advances in technology have transformed the Natural Language Technology (NLT) landscape, specifically, the use of transformers to build language models such as BERT and GPT3. Furthermore, it has been shown that the quality and the domain-specificity of input corpus to language models can improve downstream application results. However, Earth science research has minimal efforts focused on building and using a domain-specific language model. We utilize a transfer learning solution that uses an existing language model trained for general science (SciBERT) and fine-tune it using abstracts and full text extracted from various Earth science journals to create BERT-E (BERT for Earth Science). The training process utilized the input of 270k+ Earth science articles with almost 6 million paragraphs. We used Masked Language Modeling (MLM) to train the transformer model. MLM works by masking random words in the paragraph and optimizing the model for predicting the right masked word. BERT-E was evaluated by performing a downstream keyword classification task, and the performance was compared against classification results using the original SciBERT Language Model. The SciBERT-based model attained an accuracy of 89.99, whereas the BERT-E-based model attained an accuracy of 92.18, showing an improvement in overall performance.We investigate employing language models to provide new semantic search capabilities for unstructured text such as papers. This search capability requires utilizing a knowledge graph generated from Earth science corpora with a language model and convolutions to surface latent and related sentences for a natural language query. The sentences in the papers are modeled in the graph as nodes, and these nodes are connected through entities. The language model is used to give sentences a numeric representation. Graph convolutions are then applied to sentence embeddings to obtain a vector representation of the sentence along with combined representation of the surrounding graph structure. This approach utilizes both the power of adjacency inherently encoded in graph structures and latent knowledge captured in the language model. Our initial proof of concept prototype used SIMCSE training algorithm (and the tinyBERT architecture) as the embedding model. This framework has demonstrated an improved ability to surface relevant, latent information based on the input query. We plan to show new results using the domain-specific BERT-E model.
- Open Access
- 26 February 2021
- Published by Wiley in ESS Open Archive
Semantic Segmentation with Deep Convolutional Neural Networks for Automated Dust Detection in Goes-R Satellite Imagery
Airborne dust, including Dust storms and weaker dust traces, can have deleterious and hazardous effects on human health, agriculture, solar power generation, and aviation. Although earth observing satellites are extremely useful in monitoring dust using visible and infrared imagery, dust is often difficult to visually identify in single band imagery due to its similarities to clouds, smoke, and underlying surfaces. Furthermore, night-time dust detection is a particularly difficult problem, since radiative properties of dust mimic those of the cooling, underlying surface. The creation of false-color red-green-blue (RGB) composite imagery, specifically the EUMETSAT Dust RGB, was designed to enhance dust detection through the combination of single bands and band differences into a single composite image. However, dust is still often difficult to identify in night-time imagery even by experts. We developed a Deep Learning, UNET image segmentation model to identify airborne dust at night leveraging six GOES-16 infrared bands, with a focus on infrared and water vapor bands.The UNET model architecture is an encoder-decoder Convolutional Neural Network that does not require large amounts of training data, localizes and contextualized image data for precise segmentation, and provides fast training time for high accuracy pixel level prediction. This presentation highlights collection of the training database, development of the model, and preliminary model validation. With further model development, validation, and testing in a real-time context, probability-based dust prediction could alert weather forecasters, emergency managers, and citizens to the location and extent of impending dust storms.
- 26 February 2021
- Published by Wiley in ESS Open Archive
Open AccessSemantic Segmentation with Deep Convolutional Neural Networks for Automated Dust Detection in Goes-R Satellite Imagery
Airborne dust, including Dust storms and weaker dust traces, can have deleterious and hazardous effects on human health, agriculture, solar power generation, and aviation. Although earth observing satellites are extremely useful in monitoring dust using visible and infrared imagery, dust is often difficult to visually identify in single band imagery due to its similarities to clouds, smoke, and underlying surfaces. Furthermore, night-time dust detection is a particularly difficult problem, since radiative properties of dust mimic those of the cooling, underlying surface. The creation of false-color red-green-blue (RGB) composite imagery, specifically the EUMETSAT Dust RGB, was designed to enhance dust detection through the combination of single bands and band differences into a single composite image. However, dust is still often difficult to identify in night-time imagery even by experts. We developed a Deep Learning, UNET image segmentation model to identify airborne dust at night leveraging six GOES-16 infrared bands, with a focus on infrared and water vapor bands.The UNET model architecture is an encoder-decoder Convolutional Neural Network that does not require large amounts of training data, localizes and contextualized image data for precise segmentation, and provides fast training time for high accuracy pixel level prediction. This presentation highlights collection of the training database, development of the model, and preliminary model validation. With further model development, validation, and testing in a real-time context, probability-based dust prediction could alert weather forecasters, emergency managers, and citizens to the location and extent of impending dust storms.
- 18 January 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Radio Frequency Identification
- Vol. 5 (2) , 113-127
Exploiting Spatio-Temporal Properties of I/Q Signal Data Using 3D Convolution for RF Transmitter Identification
In this work we implement an automatic emitter identification system using three dimensional convolutional neural networks (3D-CNN). We show that it is possible to leverage the short term spatio-temporal properties of the raw I/Q signal data using 3D-CNN and apply this to the task of transmitter identification. The time series of raw I/Q signal data exhibits a “helical” structure which encodes the intrinsic properties of the emitter as short term variations. These features are exploited by our system to create unique “fingerprints” for the emitters, which are finally used for the task of identification. Our system is end-to-end, works with the raw I/Q signal data and is able to automatically discriminate between different transmitters without using a recurrent structure, thus reducing the complexity of the system. We experimented with four novel RF transmitter datasets that we curated in an uncalibrated and uncontrolled indoor environment. The first dataset consists of four identical transmitters, the second eight identical transmitters, the third four heterogeneous transmitters and the last eight heterogeneous transmitters. We use Ettus Research USRP Software Defined Radios (USRP B210, B200, X310), ADLM Pluto and BladeRF for transmission and collect the raw signal data using a RTL Software Defined Radio (RTL SDR) receiver dongle, from each of the four and eight emitters at a time. We use the data to train and test the 3D convolutional neural network for discerning intrinsic transmitter characteristics for the task of classification. Through our methods and implementation, we show that we can identify the transmitters with ~99 % accuracy.
- 18 January 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Radio Frequency Identification
- Vol. 5 (2) , 113-127
Exploiting Spatio-Temporal Properties of I/Q Signal Data Using 3D Convolution for RF Transmitter Identification
In this work we implement an automatic emitter identification system using three dimensional convolutional neural networks (3D-CNN). We show that it is possible to leverage the short term spatio-temporal properties of the raw I/Q signal data using 3D-CNN and apply this to the task of transmitter identification. The time series of raw I/Q signal data exhibits a “helical” structure which encodes the intrinsic properties of the emitter as short term variations. These features are exploited by our system to create unique “fingerprints” for the emitters, which are finally used for the task of identification. Our system is end-to-end, works with the raw I/Q signal data and is able to automatically discriminate between different transmitters without using a recurrent structure, thus reducing the complexity of the system. We experimented with four novel RF transmitter datasets that we curated in an uncalibrated and uncontrolled indoor environment. The first dataset consists of four identical transmitters, the second eight identical transmitters, the third four heterogeneous transmitters and the last eight heterogeneous transmitters. We use Ettus Research USRP Software Defined Radios (USRP B210, B200, X310), ADLM Pluto and BladeRF for transmission and collect the raw signal data using a RTL Software Defined Radio (RTL SDR) receiver dongle, from each of the four and eight emitters at a time. We use the data to train and test the 3D convolutional neural network for discerning intrinsic transmitter characteristics for the task of classification. Through our methods and implementation, we show that we can identify the transmitters with ~99 % accuracy.