Peer-Reviewed Publications

Early Identification of Deforestation using Anomaly Detection

Authors: N. Wijesinghe, R. Perera, N. Sellahewa, P. D. Talagala
Published in: 2023 8th International Conference on Information Technology Research (ICITR), Colombo, Sri Lanka, 2023, pp. 1-6
DOI: 10.1109/ICITR61062.2023.10382919

Abstract:

Research involving anomaly detection in image streams has seen growth through the years, given the proliferation of high-quality image data in various applications. One such application that is in urgent need of attention is deforestation. Detecting anomalies in this context, however, remains challenging due to the irregular and low-probability nature of deforestation events. This study introduces two anomaly detection frameworks utilizing machine learning and deep learning for the early detection of deforestation activities in image streams. Furthermore, Explainable AI was used to explain the black box models of the deep learning-based anomaly detection framework. The class imbalance problem, the inter-dependency between the images with time, the lack of available labelled images, a data-driven anomalous threshold, and the trade-off of accuracy while increasing interpretability in the black box optimization methods are some key aspects considered in the model-building process. Our novel framework for anomaly detection in image streams underwent rigorous evaluation using a range of datasets that included synthetic and real-world data, notably datasets related to Amazon’s forest coverage. The objective of this evaluation was to detect occurrences of deforestation in the Amazon. Several metrics were used to evaluate the performance of the proposed framework.

Keywords: Deep learning, Deforestation, Closed box, Feature extraction, Anomaly detection

🔗 View on IEEE Xplore

Cross-vit: Cross-attention vision transformer for image duplicate detection

Authors: M. D. N. Chandrasiri and P. D. Talagala
Published in: 2023 8th International Conference on Information Technology Research (ICITR), Colombo, Sri Lanka, 2023, pp. 1-6
DOI: 10.1109/ICITR61062.2023.10382916

Abstract:

Research involving anomaly detection in image streams has seen growth through the years, given the proliferation of high-quality image data in various applications. One such application that is in urgent need of attention is deforestation. Detecting anomalies in this context, however, remains challenging due to the irregular and low-probability nature of deforestation events. This study introduces two anomaly detection frameworks utilizing machine learning and deep learning for the early detection of deforestation activities in image streams. Furthermore, Explainable AI was used to explain the black box models of the deep learning-based anomaly detection framework. The class imbalance problem, the inter-dependency between the images with time, the lack of available labelled images, a data-driven anomalous threshold, and the trade-off of accuracy while increasing interpretability in the black box optimization methods are some key aspects considered in the model-building process. Our novel framework for anomaly detection in image streams underwent rigorous evaluation using a range of datasets that included synthetic and real-world data, notably datasets related to Amazon’s forest coverage. The objective of this evaluation was to detect occurrences of deforestation in the Amazon. Several metrics were used to evaluate the performance of the proposed framework.

Keywords: Duplicate Image Detection, Vision Transformers, Attention

🔗 View on IEEE Xplore

From Crisis to Opportunity: A Google Trends Analysis of Global Interest in Distance Education Tools During and Post the COVID-19 Pandemic

Authors: Priyanga Dilini Talagala, Thiyanga S. Talagala
Published in: 6th International Conference on Advanced Research Methods and Analytics (CARMA 2024), Valencia, 26-28 June 2024
DOI: 10.4995/CARMA2024.2024.17804

Abstract:

This study investigated the impact of COVID-19 on global attention towards different distance education tools. We used Google Trend search queries as a proxy to quantify the popularity and public interest in different distance education solutions under 11 subsegments, which include collaboration platforms, online proctoring, and resources for psychosocial support. The study employs both visual and analytical approaches to analyse global web search queries during and post the COVID-19 pandemic. Through cross-correlation analysis and dynamic time-warping analysis, the study confirms the contemporaneous and lead-lag relationships between COVID-19 and distance education-related search terms. Furthermore, the study highlights the critical role of psychosocial support in promoting the well-being of students and teachers during a pandemic. The study emphasizes the importance of Google footprint analysis in determining the most popular online education resources designed for different educational goals. This feature allows educators to gain insight into prominent distant education options, boosting their online teaching.

Keywords: Online Learning; Online Teaching; Distance Education Solutions; COVID-19 Pandemic; Google Trend Search Queries; Psychosocial Support in Education

🔗 View Conference Paper

Generalized Meta Framework for Forecasting

Authors: Theepana Govintharajah, Pavadaran Pathmaranjan, Gowsigan Kanagalingam, Priyanga Dilini Talagala

Published in: 2024 9th International Conference on Information Technology Research (ICITR), Colombo, Sri Lanka, 05-06 December 2024
DOI: 10.1109/ICITR64794.2024.10857715

Abstract:

Forecasting tabular time series data has become a challenging task as the time series data has its own unique patterns, and therefore identifying the most suitable modeling approach for a given dataset requires additional investigations and expert knowledge. In this study, we propose a novel meta framework that utilizes an ensemble approach, combining the models with a high level of performance and efficiency for a given dataset with the aim of proposing a more generalized framework for time series forecasting. In this proposed approach, we use different models from a large pool of candidate models that have the ability to capture unique time series characteristics available in various time series datasets. This approach allows us to get a more robust and generalized framework. By using a variety of forecasting models, including statistics-based prediction models, machine learning-based prediction models, deep learning-based prediction models, and generative modeling, our novel approach ensures broad applicability across various datasets from different application domains. Meta features that describe the structure, complexity, and time series patterns available in a dataset are used to determine the optimal ensemble from a large pool of candidate models for the final prediction process. A hybrid architecture incorporates generative models into the pool, and a stacking approach integrates the predictions of the ensemble’s member models.

Keywords: Time Series Forecasting, Meta-Learning, Ensemble Modeling, Generative Models, Deep Learning, Machine Learning

🔗 View on IEEE Xplore

Early Disease Outbreak Detection in Spatio-Temporal Data Using Predictive Modeling and Extreme Value Theory

Authors: E.G.M.A. Senevirathne, Priyanga D. Talagala

Published in: 2024 9th International Conference on Information Technology Research (ICITR), Colombo, Sri Lanka, 05-06 December 2024
DOI: 10.1109/ICITR64794.2024.10857720

Abstract:

Early detection of outbreaks is crucial for reducing their impact on public health. Static manual thresholds have been used for traditional detection methods, which fail to capture extreme events in dynamic transmission patterns. The aim of this study is to introduce a generalized framework that integrates feature engineering, predictive modeling, and Extreme Value Theory (EVT) for dynamic thresholding in spatio-temporal data. This framework is capable of adapting to different diseases and regions, enabling more accurate outbreak detection across different datasets. Applied to dengue and COVID-19 case data, the proposed method outperformed traditional approaches by achieving higher accuracy, precision, and F1 scores. The EVT-based method provides a more reliable solution for identifying outbreaks in irregularly distributed data, enhancing public health response capabilities.

Keywords: Disease Outbreak Detection, Spatio-Temporal Data, Predictive Modeling, Extreme Value Theory, Public Health

🔗 View on IEEE Xplore

Enhancing Demand Forecasting in Food Manufacturing: Hierarchical Analysis of Aggregated and Individual Models

Authors: Achala Hasini Perera, Priyanga Dilini Talagala, H. Niles Perera, Amila Thibbotuwawa

Published in: 2024 9th International Conference on Information Technology Research (ICITR), Colombo, Sri Lanka, 05-06 December 2024
DOI: 10.1109/ICITR64794.2024.10857772

Abstract:

This study focuses on production planning in the food manufacturing sector using hierarchical forecasting. The selected case for the focal study represents food products with a common main ingredient used in manufacturing. We employ two scenarios: 1) forecasting aggregated total sales for all products, and 2) forecasting sales for each product separately to calculate the total requirement. We employed three statistical models: autoregressive integrated moving average (ARIMA), exponential smoothing (ETS), and Prophet, and five machine learning models such as linear regression (LR), k-nearest neighbors (KNN), support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGBoost). The key findings highlight that forecasting aggregated total sales for common ingredients outperformed forecasting for each product individually. Further, island-level forecasts were more accurate than district- and distribution-center-level forecasts. XGBoost performed as the best forecasting model, and MinT outperformed as the best reconciliation approach. This study contributes to supply chain strategies when products have common ingredients, optimizing resource allocation and production planning, enhancing operational efficiency in food manufacturing.

Keywords: Demand Forecasting, Food Manufacturing, Hierarchical Forecasting, Machine Learning, Time Series, ARIMA, ETS, Prophet, XGBoost, MinT

🔗 View on IEEE Xplore

Improving Class Imbalance in the Classification of Multi-Dimensional Data: Interpretable Model Design and Evaluation

Authors: Gayathri Sivakumar, Chambavy Balasundaram, Vithursan Thevendran, Priyanga Dilini Talagala

Published in: 2024 9th International Conference on Information Technology Research (ICITR), Colombo, Sri Lanka, 05-06 December 2024
DOI: 10.1109/ICITR64794.2024.10857748

Abstract:

This study presents a hybrid approach that combines deep learning techniques with conventional machine learning techniques to address the class imbalance in the classification of multi-dimensional data. The resulting framework incorporates SHapley Additive exPlanations (SHAP) to evaluate the model predictions based on domain knowledge. It combines Conditional Generative Adversarial Networks (CGANs), Self-Supervised Clustered GANs (SSCGANs), and Variational Autoencoders (VAEs) for the generation of improved synthetic data. This method ensures that model decisions are based on domain-specific knowledge while enabling efficient computation of SHAP values by approximation of complex classifiers using surrogate models. Evaluations show that the suggested method overcomes the shortcomings of current techniques in high-stakes domains and improves classification performance and transparency.

Keywords: Class Imbalance, Multi-Dimensional Data, SHAP, CGAN, SSCGAN, VAE, Explainable AI, Machine Learning, Deep Learning

🔗 View on IEEE Xplore

Working Papers & Preprints

1. Chandeepa Pathirana and P. D. Talagala, “DHybrid Feature-Hash Module For Image Duplicate Detection,” 2025 10th International Conference on Information Technology Research (ICITR), Colombo, Sri Lanka, 2025, (Under Review)

2. M.H.D.B Gajakum, M. D. N. Chandrasiri and P. D. Talagala, “Dynamic Cluster Specific Ensembles with Explainable AI for Flood Hazard Prediction,” 2025 10th International Conference on Information Technology Research (ICITR), Colombo, Sri Lanka, 2025, (Under Review)

3. Wijesinghe, N., Perera, R., Sellahewa, N., & Talagala, P. D. (2024). Anomaly Detection in Image Streams Using Explainable AI (Working Paper). Faculty of Information Technology, University of Moratuwa, Sri Lanka.

Software & Tools

Talagala, P. D. (2024). clap: Detecting class overlapping regions in multidimensional data (R package version 0.1.0). https://doi.org/10.32614/CRAN.package.clap

You can install the clap R package in R using either CRAN or GitHub:

# Install from CRAN
install.packages("clap")

# Or install the development version from GitHub
# install.packages("devtools") # if not already installed
devtools::install_github("pridiltal/clap")

P. D. Talagala, R. J. Hyndman, and G. Romano, “AnomalyDetection: CRAN Task View: Anomaly Detection,” GitHub repository, version 2025-10-21, accessed Nov. 15, 2025. [Online]. Available: https://github.com/cran-task-views/AnomalyDetection/

You can install the AnomalyDetection CRAN Task View in R with a single line using the ctv R package:

# Install the ctv package if you don't have it
install.packages("ctv")

# Install the AnomalyDetection Task View
install.views("AnomalyDetection")