My Research

Sponsored R&D Projects

About

My research interests lie primarily in the fields of remote sensing, pattern recognition and machine learning. During my doctoral studies, I have investigated various neural approaches for domain adaptation for land-cover classification using remotely sensed satellite images. Domain adaptation is a well known problem in the machine learning community which deals with exploring unknown domain given abundant available knowledge from a different domain. In pattern recognition, the two domains are said to be different if the sample distribution of classes across the two domains have discrepancies. In such situations, a machine learning based classifier trained on available labelled information from a particular domain fails to predict the class labels for samples from the other domain. A variety of neural network based techniques is currently being worked out towards facilitating adaptability in land-cover classification.

In addition to this, I have also investigated the power efficiency of wireless sensor routing protocols for disaster mitigation in fulfillment to my master degree research project. In this project, the integrity of well known routing protocols for low-power communication of wireless sensors have been re-investigated under scenarios of natural disaster. A hybrid protocol having a better power efficiency has been proposed after a comparative study of the routing protocols existent in literature. Furthermore, the proposed protocol has later been extended to remain operational in situations of natural catastrophe. 

Moreover, a optical character recognition system has also been investigated by me during my under-graduate course where ideas have been developed for character recognition using statistical machine learning techniques. It is during this graduation project when I was first motivated to the research related to the learning systems. The recognition has been carried out by exploring statistical features of the printed alphabets like skew, kurtosis and symmetry.

Some details can be found here

PhD Thesis: Neural approaches towards adaptive land-cover classification using remotely sensed images (June 2015 – August 2020)

Land-cover classification (LCC) deals with the identification of land-cover classes like soil, urban, water, crop, snow, etc. If carried out in a timely manner, it can be useful in important Earth observation applications like deforestation control, land-change detection, thermal mapping, navigation, sea-ice, and agriculture monitoring. Traditionally, LCC has been carried out manually by field surveyors and the process is inaccurate, time-consuming and costly in terms of human effort. As a consequence, the manual approaches have soon been replaced by intelligent automatic LCC techniques aided by machine learning (ML) technologies. An automatic LCC technique generates a land-cover map based on the class predictions from a ML-based classifier trained over some of the labelled samples corresponding to the pixels in the image. However, the conventional LCC techniques fail when the target image has a different sample distribution than that of the labelled examples available in the training set. To further reduce the human effort in a fresh collection of labelled samples from each of the unexplored regions, some adaptive LCC techniques have been worked out in this thesis by reengineering the already available training set to suit target classification. The purpose here is to develop artificial neural network-based solutions to this domain adaptation (DA) problem in LCC by integrating unsupervised, semi-supervised and active learning in a hybrid framework. At the onset, an unsupervised DA technique has been explored using ensemble of intermediate source-target feature maps computed by auto-encoder and restricted Boltzmann machine neural networks. However to improve the target accuracy, some amount of (semi)labelled target information has been incorporated in the training set through semi-supervised and active learning techniques, respectively. Further, a solution to the heterogeneous DA problem has been investigated in the latter where the target (new) image has disparate set of classes with respect to the initially available training set. Finally, the capabilities of a stacked auto-encoder have been exploited for multi-levelled, weighted and granular transformation of samples across source-target sub-groups. To assess the performance of the proposed methodologies, experiments have been conducted on the samples extracted from different multispectral satellite images captured over various regions of India. Through an analysis of the obtained results, it can be concluded that the proposed DA schemes have an edge over the state-of-the art DA schemes in accurate prediction of the target classes.

Collaborative works related to PhD

Handling the class-imbalance problem in land-cover classification (2018-19): This work has been carried out in collaboration with Prof. B. B. Chaudhuri, Indian Statistical Institute, Kolkata where classification has been carried out in situations where labelled patterns from some of the classes out-number others in the training set. This problem has been tackled using a bagging-based semi-supervised approach and a novel imbalance aware combination rule for obtaining a unanimous decision from a fusion of classifiers.

Semi-supervised two-level fusion based auto-encoded approach for low-cost adaptation of remotely sensed images (2017-18):

In this domain adaptation technique, an auto-encoding system has been investigated then to transform the samples from the paired target cluster in terms of their mapped source cluster. Thereafter, the unique target domain distribution is modelled using very few (i.e., low cost) actively collected target patterns from each of the alien clusters. This work has been carried out in collaboration with Prof. Farid Melgani, University of Trento, Italy.

Overview of the other major projects undertaken

Rapid screening technology for identification of COVID-19 carriers using radiological images (2020): An automatic image processing based smart system is being developed for identification of possible COVID-19 carriers from X-ray and computed tomography (CT) scans of human lungs. This is necessary for identification and isolation of people with COVID-19 infections in places of mass gathering. This first-look diagnosis system will automate the COVID-19 screening process replacing the inconclusive thermal scan-based (manual) procedure being used currently. After submission of PhD thesis, I have teamed up to work on this problem and this idea has been selected among the top 200 among 2500 plus entries in COVID-Samadhan competition hosted by MHRD, Govt. of India.

Power efficiency of wireless sensor routing protocols for disaster mitigation (2013-14): In this project, the integrity of well-known routing protocols for low-power communication of wireless sensors have been reinvestigated under scenarios of natural disaster. A hybrid protocol having a better power efficiency has been proposed after a comparative study of the routing protocols existent in literature. Furthermore, the proposed protocol has later been extended to remain operational in situations of natural catastrophe. This project was done in fulfilment to the master degree.

Optical Character Recognition (2011-12): Here, ideas have been developed for character recognition using statistical machine learning techniques. It is during this graduation project when I was first motivated to the research related to the learning systems. The recognition has been carried out by exploring statistical features of the printed alphabets like skew, kurtosis and symmetry.