Uttam Kumar
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Uttam Kumar has a total of 58 co-authors. Here are the top co-authors by number of shared publications:
- Thejaswini Ramachandra
Energy & Wetlands Research Group, Environmental Information System, CES TE 15, Center for Ecological Sciences [CES], Indian Institute of Science, Bangalore, India
10 shared publications Chiranjit Mukhopadhyay
Department of Management Studies, Indian Institute of Science, Bangalore, India
5 shared publicationsS. Kumar Raja
5 shared publicationsCristina Milesi
NASA Ames Research Ctr. (United States)
4 shared publicationsRamakrishna R. Nemani
NASA Ames Research Ctr. (United States)
4 shared publications- Ramakrishna R. Nemani
NASA Ames Research Center Moffett Field CA USA
4 shared publications
Latest Publications
- Open Access
- 10 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 5258-5262
Caption, Create, Continue: Continual Learning with Pre-trained Generative Vision-Language Models
- 10 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 5258-5262
Open AccessCaption, Create, Continue: Continual Learning with Pre-trained Generative Vision-Language Models
- Open Access
- 12 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Spatiotemporal Forecasting of Traffic Flow using Wavelet-based Temporal Attention
Spatiotemporal forecasting of traffic flow data represents a typical problem impacting urban traffic management systems. In general, spatiotemporal forecasting problems involve complex interactions, nonlinearities, and long-range dependencies due to the interwoven nature of the temporal and spatial dimensions. Due to this, traditional statistical and machine learning methods cannot adequately handle both the temporal and spatial dependencies in traffic flow datasets. A prevalent approach in the field combines graph convolutional networks and multi-head attention mechanisms for spatiotemporal processing of traffic flow data. This paper proposes a wavelet-based temporal attention model, namely a wavelet-based dynamic spatiotemporal aware graph neural networks (W-DSTAGNN), for tackling the traffic forecasting problem. A key feature of W-DSTAGNN is the use of wavelets that can effectively separate noise and capture multi-resolution temporal patterns. Wavelet decomposition can help by decomposing the signal into components that can be analyzed independently, reducing the impact of non-stationarity and handling long-range dependencies of traffic flow datasets. These enable our proposal to generate more robust and better accurate forecasts of complex and dynamic traffic dependencies than commonly used spatiotemporal deep learning models. Benchmark experiments using three popularly used statistical metrics confirm that our proposal efficiently captures spatiotemporal correlations and outperforms eleven state-of-the-art models (including both temporal and spatiotemporal benchmarks) on three publicly available traffic datasets. Our proposed ensemble method can better handle dynamic temporal and spatial dependencies and make reliable long-term forecasts. In addition to point forecasts, our proposed model can generate interval forecasts that significantly enhance probabilistic forecasting of traffic datasets.
- 12 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Open AccessSpatiotemporal Forecasting of Traffic Flow using Wavelet-based Temporal Attention
Spatiotemporal forecasting of traffic flow data represents a typical problem impacting urban traffic management systems. In general, spatiotemporal forecasting problems involve complex interactions, nonlinearities, and long-range dependencies due to the interwoven nature of the temporal and spatial dimensions. Due to this, traditional statistical and machine learning methods cannot adequately handle both the temporal and spatial dependencies in traffic flow datasets. A prevalent approach in the field combines graph convolutional networks and multi-head attention mechanisms for spatiotemporal processing of traffic flow data. This paper proposes a wavelet-based temporal attention model, namely a wavelet-based dynamic spatiotemporal aware graph neural networks (W-DSTAGNN), for tackling the traffic forecasting problem. A key feature of W-DSTAGNN is the use of wavelets that can effectively separate noise and capture multi-resolution temporal patterns. Wavelet decomposition can help by decomposing the signal into components that can be analyzed independently, reducing the impact of non-stationarity and handling long-range dependencies of traffic flow datasets. These enable our proposal to generate more robust and better accurate forecasts of complex and dynamic traffic dependencies than commonly used spatiotemporal deep learning models. Benchmark experiments using three popularly used statistical metrics confirm that our proposal efficiently captures spatiotemporal correlations and outperforms eleven state-of-the-art models (including both temporal and spatiotemporal benchmarks) on three publicly available traffic datasets. Our proposed ensemble method can better handle dynamic temporal and spatial dependencies and make reliable long-term forecasts. In addition to point forecasts, our proposed model can generate interval forecasts that significantly enhance probabilistic forecasting of traffic datasets.
- Open Access
- 29 January 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Searching for a new probability distribution for modeling non-scale-free heavy-tailed real-world networks
In this study, we consider large-scale network data sets from different disciplines, namely social networks, collaboration networks, web graphs, citation networks, biological networks, product co-purchasing networks, temporal networks, communication networks, ground-truth networks, and brain networks. We study several individual data sets from each discipline. These data sets are publicly available at http://snap.stanford.edu/data/index.html Usage of the proposed probability models for the paper "Searching for a new probability distribution for modeling non-scale-free heavy-tailed real-world networks" is available at https://github.com/tanujit123/nonscalefree
- 29 January 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Open AccessSearching for a new probability distribution for modeling non-scale-free heavy-tailed real-world networks
In this study, we consider large-scale network data sets from different disciplines, namely social networks, collaboration networks, web graphs, citation networks, biological networks, product co-purchasing networks, temporal networks, communication networks, ground-truth networks, and brain networks. We study several individual data sets from each discipline. These data sets are publicly available at http://snap.stanford.edu/data/index.html Usage of the proposed probability models for the paper "Searching for a new probability distribution for modeling non-scale-free heavy-tailed real-world networks" is available at https://github.com/tanujit123/nonscalefree
- 1 November 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1884-1889
Automatic feature selection for Landsat-8 and Sentinel-2 image co-registration using SURFFANN
Remote sensing data acquired from Landsat series of satellites having multispectral bands with a moderate spatial resolution of 30 m are available from 1973 and Sentine1-2 data at 10 m spatial resolution are available from 2015 onwards with a regular temporal global coverage. The availability of Landsat-8 OLI (Operational Land Imager) and Sentine1-2 MSI (Multispectral Instrument) imagery have offered opportunities to build a wide range of spatio-temporal applications such as monitoring the Earth's surface frequently based on the combined data sets. However, the standard georeferenced Landsat-8 OLI and Sentine1-2 MSI data products are misaligned, therefore the two data sets must be co-registered at a pixel level before being used together for any study. In this paper, we present a new approach called SURF-FANN (Speeded-Up Robust Features-Fast Approximate Nearest Neighbour) for feature selection and matching between Landsat and Sentinel data. These features (tie-points) are then used for automatic co-registration of Sentinel data with reference to Landsat data. Further, geocoding is performed using conformal, affine and projective transformations followed by resampling using bilinear interpolation, nearest neighbour and cubic convolution. The results were compared with SIFT (Scale Invariant Feature Transform)-FANN which is a popular algorithm for automatic feature extraction and selection. It was found that SURF is computationally faster but has higher RMSE compared to SIFT. However, the feature selection and matching by combining SURF and SIFT with FANN, followed by projective transformation and bilinear interpolation rendered the lowest RMSE of 0.14 compared to using either SIFT or SURF as a stand-alone method.
- 1 November 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1884-1889
Automatic feature selection for Landsat-8 and Sentinel-2 image co-registration using SURFFANN
Remote sensing data acquired from Landsat series of satellites having multispectral bands with a moderate spatial resolution of 30 m are available from 1973 and Sentine1-2 data at 10 m spatial resolution are available from 2015 onwards with a regular temporal global coverage. The availability of Landsat-8 OLI (Operational Land Imager) and Sentine1-2 MSI (Multispectral Instrument) imagery have offered opportunities to build a wide range of spatio-temporal applications such as monitoring the Earth's surface frequently based on the combined data sets. However, the standard georeferenced Landsat-8 OLI and Sentine1-2 MSI data products are misaligned, therefore the two data sets must be co-registered at a pixel level before being used together for any study. In this paper, we present a new approach called SURF-FANN (Speeded-Up Robust Features-Fast Approximate Nearest Neighbour) for feature selection and matching between Landsat and Sentinel data. These features (tie-points) are then used for automatic co-registration of Sentinel data with reference to Landsat data. Further, geocoding is performed using conformal, affine and projective transformations followed by resampling using bilinear interpolation, nearest neighbour and cubic convolution. The results were compared with SIFT (Scale Invariant Feature Transform)-FANN which is a popular algorithm for automatic feature extraction and selection. It was found that SURF is computationally faster but has higher RMSE compared to SIFT. However, the feature selection and matching by combining SURF and SIFT with FANN, followed by projective transformation and bilinear interpolation rendered the lowest RMSE of 0.14 compared to using either SIFT or SURF as a stand-alone method.
- Open Access
- 29 October 2017
- Published by MDPI AG in Remote Sensing
- Vol. 9  (11) , 1105
Exploring Subpixel Learning Algorithms for Estimating Global Land Cover Fractions from Satellite Data Using High Performance Computing
Land cover (LC) refers to the physical and biological cover present over the Earth’s surface in terms of the natural environment such as vegetation, water, bare soil, etc. Most LC features occur at finer spatial scales compared to the resolution of primary remote sensing satellites. Therefore, observed data are a mixture of spectral signatures of two or more LC features resulting in mixed pixels. One solution to the mixed pixel problem is the use of subpixel learning algorithms to disintegrate the pixel spectrum into its constituent spectra. Despite the popularity and existing research conducted on the topic, the most appropriate approach is still under debate. As an attempt to address this question, we compared the performance of several subpixel learning algorithms based on least squares, sparse regression, signal–subspace and geometrical methods. Analysis of the results obtained through computer-simulated and Landsat data indicated that fully constrained least squares (FCLS) outperformed the other techniques. Further, FCLS was used to unmix global Web-Enabled Landsat Data to obtain abundances of substrate (S), vegetation (V) and dark object (D) classes. Due to the sheer nature of data and computational needs, we leveraged the NASA Earth Exchange (NEX) high-performance computing architecture to optimize and scale our algorithm for large-scale processing. Subsequently, the S-V-D abundance maps were characterized into four classes, namely forest, farmland, water and urban areas (in conjunction with nighttime lights data) over California, USA using a random forest classifier. Validation of these LC maps with the National Land Cover Database 2011 products and North American Forest Dynamics static forest map shows a 6% improvement in unmixing-based classification relative to per-pixel classification. As such, abundance maps continue to offer a useful alternative to high-spatial-resolution classified maps for forest inventory analysis, multi-class mapping, multi-temporal trend analysis, etc.
- 29 October 2017
- Published by MDPI AG in Remote Sensing
- Vol. 9  (11) , 1105
Open AccessExploring Subpixel Learning Algorithms for Estimating Global Land Cover Fractions from Satellite Data Using High Performance Computing
Land cover (LC) refers to the physical and biological cover present over the Earth’s surface in terms of the natural environment such as vegetation, water, bare soil, etc. Most LC features occur at finer spatial scales compared to the resolution of primary remote sensing satellites. Therefore, observed data are a mixture of spectral signatures of two or more LC features resulting in mixed pixels. One solution to the mixed pixel problem is the use of subpixel learning algorithms to disintegrate the pixel spectrum into its constituent spectra. Despite the popularity and existing research conducted on the topic, the most appropriate approach is still under debate. As an attempt to address this question, we compared the performance of several subpixel learning algorithms based on least squares, sparse regression, signal–subspace and geometrical methods. Analysis of the results obtained through computer-simulated and Landsat data indicated that fully constrained least squares (FCLS) outperformed the other techniques. Further, FCLS was used to unmix global Web-Enabled Landsat Data to obtain abundances of substrate (S), vegetation (V) and dark object (D) classes. Due to the sheer nature of data and computational needs, we leveraged the NASA Earth Exchange (NEX) high-performance computing architecture to optimize and scale our algorithm for large-scale processing. Subsequently, the S-V-D abundance maps were characterized into four classes, namely forest, farmland, water and urban areas (in conjunction with nighttime lights data) over California, USA using a random forest classifier. Validation of these LC maps with the National Land Cover Database 2011 products and North American Forest Dynamics static forest map shows a 6% improvement in unmixing-based classification relative to per-pixel classification. As such, abundance maps continue to offer a useful alternative to high-spatial-resolution classified maps for forest inventory analysis, multi-class mapping, multi-temporal trend analysis, etc.
- 1 August 2017
- Published by Taylor & Francis
- p. 113-130
Deep Learning for Very High-Resolution Imagery Classification
This chapter demonstrates the use of a scalable machine learning algorithm using airborne imagery data acquired by the National Agriculture Imagery Program (NAIP) for the Continental United States (CONUS) at an optimal spatial resolution of 1 m. It focuses on leveraging the power of deep belief networks (DBNs) for the representation and classification of aerial imagery data acquired at very high spatial resolutions of the order of 1 m. The chapter provides a detailed description of the unsupervised segmentation phase using the SRM algorithm. It enumerates the key components of the feature extraction step and the learning phase based on DBN. The chapter illustrates the details of the CRF algorithm and the online update procedure for the training data. It provides an overview of the NASA Earth Exchange high-performance computing (NEX HPC) architecture. The chapter contains results and comparative studies with National Land Cover Database (NLCD), and light detection and ranging (LiDAR). It also provides details of the AWS infrastructure.
- 1 August 2017
- Published by Taylor & Francis
- p. 113-130
Deep Learning for Very High-Resolution Imagery Classification
This chapter demonstrates the use of a scalable machine learning algorithm using airborne imagery data acquired by the National Agriculture Imagery Program (NAIP) for the Continental United States (CONUS) at an optimal spatial resolution of 1 m. It focuses on leveraging the power of deep belief networks (DBNs) for the representation and classification of aerial imagery data acquired at very high spatial resolutions of the order of 1 m. The chapter provides a detailed description of the unsupervised segmentation phase using the SRM algorithm. It enumerates the key components of the feature extraction step and the learning phase based on DBN. The chapter illustrates the details of the CRF algorithm and the online update procedure for the training data. It provides an overview of the NASA Earth Exchange high-performance computing (NEX HPC) architecture. The chapter contains results and comparative studies with National Land Cover Database (NLCD), and light detection and ranging (LiDAR). It also provides details of the AWS infrastructure.
- 4 July 2017
- Published by Springer Nature in Journal of the Indian Society of Remote Sensing
- Vol. 46  (3) , 407-422
Examining the Effect of Ancillary and Derived Geographical Data on Improvement of Per-Pixel Classification Accuracy of Different Landscapes
- 4 July 2017
- Published by Springer Nature in Journal of the Indian Society of Remote Sensing
- Vol. 46  (3) , 407-422
Examining the Effect of Ancillary and Derived Geographical Data on Improvement of Per-Pixel Classification Accuracy of Different Landscapes
- 4 September 2015
- Published by SPIE-Intl Soc Optical Eng
- Vol. 9610, 96100B-96100B-15
Land cover fraction estimation with global endmembers using collaborative SUnSAL
- 4 September 2015
- Published by SPIE-Intl Soc Optical Eng
- Vol. 9610, 96100B-96100B-15
Land cover fraction estimation with global endmembers using collaborative SUnSAL
- 1 August 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 3753, 244-251
Simplex Projection for Land Cover Information Mining from Landsat-5 TM Data
In this paper we explore the efficacy of simplex projection for land cover (LC) information mining. LC is the observed biophysical cover on the surface of the Earth and describes how much of a region is covered by forests, wetlands, impervious surfaces, etc. LC information can be extracted by processing remotely sensed data acquired through sensors mounted either on space borne satellites or aircrafts. Since these data are a mixture of more than two LC class types, unmixing algorithms based on linear mixture model such as simplex projection, aims to resolve the different components of mixed pixels in the data. This method is based on the equivalence of the fully constrained least squares problem of projecting a point onto a simplex. The algorithm does not perform optimization and is analytical, thus reducing the computational complexity. The algorithm is tested on computer-simulated data of various signal to noise ratio and Landsat-5 TM data of an agricultural landscape and an urban scenario. The results are validated using descriptive statistics, correlation coefficient, root mean square error, probability of success and bivariate distribution function.
- 1 August 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 3753, 244-251
Simplex Projection for Land Cover Information Mining from Landsat-5 TM Data
In this paper we explore the efficacy of simplex projection for land cover (LC) information mining. LC is the observed biophysical cover on the surface of the Earth and describes how much of a region is covered by forests, wetlands, impervious surfaces, etc. LC information can be extracted by processing remotely sensed data acquired through sensors mounted either on space borne satellites or aircrafts. Since these data are a mixture of more than two LC class types, unmixing algorithms based on linear mixture model such as simplex projection, aims to resolve the different components of mixed pixels in the data. This method is based on the equivalence of the fully constrained least squares problem of projecting a point onto a simplex. The algorithm does not perform optimization and is analytical, thus reducing the computational complexity. The algorithm is tested on computer-simulated data of various signal to noise ratio and Landsat-5 TM data of an agricultural landscape and an urban scenario. The results are validated using descriptive statistics, correlation coefficient, root mean square error, probability of success and bivariate distribution function.
- 1 August 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 3, 574-581
Unconstrained Linear Spectral Mixture Models for Spatial Information Extraction: A Comparative Study
Linear spectral unmixing remains an important subject of research in image processing community. It deals with disintegrating a pixel spectrum into its constituent spectra through a mixture model assuming that observed data are linear mixtures of two or more objects, representing a mixed pixel. The linear mixture model allows a number of target materials (classes) to be present, each contributing a fraction of its unique, fixed spectrum corresponding to the area occupied by that material in a pixel. The linear model is inverted to produce estimates of those fractional abundances. The optimal solution of the mixture model can be an unconstrained solution, partially constrained solution or a fully constrained solution when the abundance non-negativity and abundance sum-to-one constraints are imposed. Although constrained algorithms are appropriate for target quantification and for estimating abundance fractions, unconstrained models are more suitable for applications seeking target detection, identification and discrimination. In this paper we perform a comparative study of five unconstrained mixture models -- unconstrained least squares, orthogonal subspace projection, singular value decomposition, sparse unmixing via variable splitting and augmented Lagrangian (SUnSAL) and SUnSAL Total Variation (SUnSAL TV). The algorithms are tested on computer-simulated data of various signal to noise ratio and Landsat-5 TM data of an agricultural landscape and an urban scenario. The results were validated using descriptive statistics, correlation coefficient, RMSE, probability of success, and bivariate distribution function.
- 1 August 2015
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 3, 574-581
Unconstrained Linear Spectral Mixture Models for Spatial Information Extraction: A Comparative Study
Linear spectral unmixing remains an important subject of research in image processing community. It deals with disintegrating a pixel spectrum into its constituent spectra through a mixture model assuming that observed data are linear mixtures of two or more objects, representing a mixed pixel. The linear mixture model allows a number of target materials (classes) to be present, each contributing a fraction of its unique, fixed spectrum corresponding to the area occupied by that material in a pixel. The linear model is inverted to produce estimates of those fractional abundances. The optimal solution of the mixture model can be an unconstrained solution, partially constrained solution or a fully constrained solution when the abundance non-negativity and abundance sum-to-one constraints are imposed. Although constrained algorithms are appropriate for target quantification and for estimating abundance fractions, unconstrained models are more suitable for applications seeking target detection, identification and discrimination. In this paper we perform a comparative study of five unconstrained mixture models -- unconstrained least squares, orthogonal subspace projection, singular value decomposition, sparse unmixing via variable splitting and augmented Lagrangian (SUnSAL) and SUnSAL Total Variation (SUnSAL TV). The algorithms are tested on computer-simulated data of various signal to noise ratio and Landsat-5 TM data of an agricultural landscape and an urban scenario. The results were validated using descriptive statistics, correlation coefficient, RMSE, probability of success, and bivariate distribution function.