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


Banage Thanne Gedara Samantha Kumara has a total of 77 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Possible Distillation for Abstractive Summarization: Reducing Hallucination in Compact LLMs

    by Vadivel Abishethvarman, Gobihanath Balasubramaniam, Senthan Prasanth, Banage T. G. S. Kumara, Banujan Kuhaneswaran

  • Open Access

    Covariance fused hybrid decomposition for probabilistic flood mapping using dual-polarized Sentinel-1 SAR using Markov Random Field optimization

    by Duminda Ranganath Welikanna, Jin Shengye, B. T. G. S. Kumara

    Accurate and scalable flood mapping using freely available synthetic aperture radar (SAR) remains challenging with dual polarization constraints and heterogeneous ground conditions under conventional intensity based approaches. This demands supplementary optimal polarimetric features to resolve the complex scattering behavior. To address this issue the study presents a covariance fused hybrid decomposition framework for probabilistic flood mapping using dual-polarized Sentinel-1(S1) single look complex (SLC) imagery. The approach betters the standard dual polarimetric analysis by combining both polarimetric and interferometric features through a multitemporal covariance matrix (CT) derived by fusing pre and post disaster VV and VH observations. This partially compensates for the absence of full polarimetric acquisitions by enabling simultaneous characterization of polarimetric scattering state and temporal coherence dynamics. Seven complementary hybrid features entropy, scattering change, coherence loss, intensity drop, α Angle, scattering diversity and double bounce change are derived through eigen value decomposition and interferometry, employing CT. They collectively encode muti dimensional response of floods in diverse landscapes. These features were integrated using a weighted linear probability model to derive continuous flood likelihood estimation. A Markov Random Field (MRF) framework was implemented to refine the probabilistic outcomes employing hybrid iterated conditional modes (ICM) and Gibbs sampling optimization. MRF ensures the spatial contextual consistency suppressing speckle induced label instability across the four class labels, low or no floods, transitional zones, high floods and undefined. High confidence flood pixels (HCFP) are introduced to estimate class conditional parameters in a fully unsupervised approach eliminating the need for training data. Validation of the results were carried out across multiple hydrological settings encompassing urban floodplains, agricultural inundation and reservoir systems particularly affected by recent cyclone ”Ditwah” event in Sri Lanka. Two independent NDWI reference data sets derived from Sentinel 2B (S2B) and UAV based multispectral imagery were employed for evaluation. Results demonstrate robust flood detection under severe class imbalance and spatially heterogeneous scattering conditions, with meaningful separability (ROC-AUC≈ 0.60 in Urban areas, F1≈0.96 in open areas) during peak flood conditions. Remaining classification errors are subjected to boundary uncertainty and temporal acquisition offsets. The framework is fully unsupervised, computationally efficient and scalable making it suitable for large area and near real time flood mapping. The study highlights the potential of covariance fusion, hybrid decomposition and MRF to unlock the enhanced flood information from dual-polarized S1 SAR data to accommodate the increasing operational demand of flood mappin

  • Investigating the Deep Learning Performance for Deforestation Mapping Using Landsat Multispectral Data

    by A.M. Chandrika Malkanthi, B.T.G.S. Kumara, D.R. Welikanna, R.M.K.T. Rathnayaka

    Deforestation is a significant threat to the sustainability of the ecosystem, leading to adverse effects such as climate change, biodiversity loss, and socio-economic consequences. Timely monitoring of forest destruction enables effective implementation of preventive mechanisms supported by law enforcement. Advancements in remote sensing, coupled with enhanced deep learning techniques, boost efficient deforestation monitoring as these technologies support real-time analysis of complex satellite images. Thus, this study aimed to develop a classification model to identify forest areas from non-forest areas using Landsat-8 data acquired for Wilpattu park, Sri Lanka, between 2015 to 2024. We explored model building using minimal input of two bands in satellite data, facilitating low resource needs. Seven deep learning models were explored, progressing from Convolution Neural Networks to Transformer-based models to build the classifier using a set of patches of size 100×100. The results were evaluated using standard metrics such as accuracy, precision, recall, F1 score, and Kappa index. We found that SegNet outperformed the remaining models with an overall accuracy of 96.36%, F1 score of 0.97, and Kappa index of 0.92, demonstrating excellent ability to distinguish the classes. However, the efficiency of the model needs further improvement. The proposed system will contribute to deforestation detection, offering a simpler model development approach with minimum input requirements. The proposed method can be adopted to other domains where the chosen band combination supports effective detection, such as water body identification.

  • Subject-Wise Impact of VAK Learning Styles on Student Personalization Using Ensemble Learning

    by Tharsan Kanagathurai, Banage T. G. S. Kumara, Vadivel Abishethvarman, Senthan Prasanth, Banujan Kuhaneswaran

  • LIME-Enhanced Sentiment Classification for UI/UX Improvement in Sri Lankan Mobile Banking

    by Nirosana Karunanithy, Senthan Prasanth, Vadivel Abishethvarman, Piumi Ishanka, Banage T. G. S. Kumara

  • Open Access

    Markov-Based Probabilistic State Estimation for Deadlock Prediction in Distributed Systems

    by Tharushi Wijethunga, Btgs Kumara

    Deadlocks are a persistent problem in distributed queueing networks due to dynamic workloads, shared resource contention, and nondeterministic execution. Conventional deadlock detection techniques are predominantly reactive and rely on deterministic rules, which limits their effectiveness in highly dynamic environments. This study proposes a Markov-based proactive deadlock detection system that estimates deadlock risk using probabilistic state modeling. The system is modeled as a stochastic process under the Markov assumption, where future system behavior depends only on the current state. System metrics such as resource utilization, traffic intensity, and queue contention are mapped to probabilistic indicators that approximate state transition likelihoods. Using Markov-based probability estimation, the likelihood of transitioning from a safe state to a deadlockprone state is continuously evaluated. These probabilistic outputs are combined into a unified deadlock risk score and integrated with machine learning classifiers to improve detection accuracy. Simulation experiments conducted under varying contention and workload conditions show that the proposed approach achieves an average detection accuracy of 91.8%, precision of 89.4%, and recall of 93.1%, while detecting deadlock-prone states 22–30% earlier than traditional reactive methods. The results demonstrate that Markov-based probabilistic state modeling effectively captures uncertainty and dynamic system behavior, enabling proactive deadlock detection and improved system reliability in distributed queueing networks.


  • Open Access

    Predictive Deadlock Prevention for Safeguarding Societal Infrastructure and Economic Stability Through Markov Models-Consensus Intelligence

    by Tharushi Wijethunga, B.T.G.S. Kumara

    Deadlocks remain a significant challenge in distributed and cloud-based computing systems, where concurrent processes compete for limited resources, potentially leading to service unavailability and cascading system failures. This issue is particularly critical in systems supporting societal infrastructure, where reliability and timely response are essential. Conventional deadlock handling techniques are largely reactive, detecting deadlocks only after system performance has already degraded. This limitation motivates the need for predictive approaches that can identify deadlock-prone states in advance. This research presents a predictive deadlock prevention framework based on probabilistic state modeling using Markov processes. System execution is represented as a sequence of state transitions derived from observed resource allocation and waiting behaviors. Transition probabilities are used to estimate the likelihood of entering deadlock-prone states, enabling early identification of high-risk conditions prior to deadlock formation. The proposed model is evaluated using experimentally generated system traces under varying levels of resource contention. Performance is assessed using classification accuracy, precision, and recall. The results show that the model achieves measurable predictive accuracy while maintaining a balanced ability to detect deadlock-prone states and limit false alarms. These findings indicate that probabilistic state-based modeling provides an interpretable and computationally lightweight foundation for proactive deadlock prevention. This work establishes a baseline predictive framework for deadlock management and highlights its potential for extension using more advanced learning techniques to improve prediction accuracy and scalability in complex cloud environments by safeguarding societal infrastructure.

  • Open Access

    Systematic Review of Graph Neural Network and Consensus Algorithm-Based Approaches for Proactive Deadlock Detection in Distributed Systems

    by W.A.T.G.R. Wijethunga, B.T.G.S. Kumara

    Deadlocks occur when a group of processes wait indefinitely for resources held by others, forming a dependency cycle that halts system progress. In distributed systems, deadlocks are particularly challenging to detect and resolve due to their complexity and asynchronous nature, leading to system inefficiencies and potential service disruptions. Traditional methods typically address deadlocks reactively, resulting in increased downtime and resource wastage. This study proposes a proactive approach that integrates Graph Neural Networks (GNNs) and consensus algorithms to detect and prevent deadlocks in distributed systems. GNNs analyze real-time system states by modeling the dependencies between processes and resource allocations, enabling early prediction of deadlock-prone situations. Consensus algorithms then coordinate distributed nodes to agree on unified prevention strategies, ensuring reliable system-wide coordination. System logs, resource allocation states, and communication traces from a simulated distributed environment and encoded into graph structures for training and evaluation of the GNN model. The experimental results demonstrate that the proposed framework achieves a highest reduction in deadlock occurrences, improved prediction accuracy, and enhanced overall system performance compared to traditional reactive methods. Furthermore, the approach shows superior scalability and adaptability to dynamic changes, reducing system downtime and manual intervention.

  • Task-Specific Knowledge Distillation for Scalable Sentiment Classification in Low-Resource Settings

    by Gobihanath Balasubramaniam, Vadivel Abishethvarman, Banage T. G. S. Kumara, Senthan Prasanth, Banujan Kuhaneswaran


  • Open Access

    LeONet: A Hybrid Deep Learning Approach for High-Precision Code Clone Detection Using Abstract Syntax Tree Features

    by Thanoshan Vijayanandan, Kuhaneswaran Banujan, Ashan Induranga, Banage T. G. S. Kumara, Kaveenga Koswattage

    Code duplication, commonly referred to as code cloning, is not inherent in software systems but arises due to various factors, such as time constraints in meeting project deadlines. These duplications, or “code clones”, complicate the program structure and increase maintenance costs. Code clones are categorized into four types: Type-1, Type-2, Type-3, and Type-4. This study aims to address the adverse effects of code clones by introducing LeONet, a hybrid Deep Learning approach that enhances the detection of code clones in software systems. The hybrid approach, LeONet, combines LeNet-5 with Oreo’s Siamese architecture. We extracted clone method pairs from the BigCloneBench Java repository. Feature extraction was performed using Abstract Syntax Trees, which are scalable and accurately represent the syntactic structure of the source code. The performance of LeONet was compared against other classifiers including ANN, LeNet-5, Oreo’s Siamese, LightGBM, XGBoost, and Decision Tree. LeONet demonstrated superior performance among the classifiers tested, achieving the highest F1 score of 98.12%. It also compared favorably against state-of-the-art approaches, indicating its effectiveness in code clone detection. The results validate the effectiveness of LeONet in detecting code clones, outperforming existing classifiers and competing closely with advanced methods. This study underscores the potential of hybrid deep learning models and feature extraction techniques in improving the accuracy of code clone detection, providing a promising direction for future research in this area.