Scholarly Articles

Permanent URI for this communityhttps://rda.sliit.lk/handle/123456789/4890

Browse

Filters

20263

Advanced Search

Filter by

Settings

Start date: 2023Subject: search.filters.subject.Deep learning

Search Results

Now showing 1 - 3 of 3
  • No Thumbnail Available
    ItemOpen Access
    SLIF-Tomato: Inverted Residual Convolutional Block Attention Module for In-field Tomato Leaf Disease Recognition
    (Springer Science and Business Media Deutschland GmbH, 2026-06-27) George, R; Thuseethan, S; Ragel, R.G; Pakeerathan, K; Sivalingam, V; Mithuran, T
    Tomato cultivation represents a critical component of nutrition, economic development, and public health, yet it is increasingly compromised by foliar diseases that diminish yield and intensify dependency on hazardous agrochemicals. Although deep learning models have demonstrated strong capabilities for automated disease recognition, existing benchmark datasets exhibit limited real-world utility, primarily due to the absence of in-field imagery and precise annotations of diseased regions. A novel attention mechanism, Inverted Residual Convolutional Block Attention Module (IR-CBAM), is proposed, combining inverted residual blocks with the CBAM Module, and is specifically tailored to address challenges posed by in-field image variability, such as complex backgrounds and inconsistent lighting. Furthermore, this study introduces SLIF-Tomato, the Sri Lankan In-Field Tomato leaf disease dataset, which is the first complete in-field dataset comprising class labels and bounding box annotations collected under diverse real-world conditions. The proposed approach achieved 99.66% and 99.91% accuracy rates on two curated versions of the SLIF-Tomato dataset. Subsequently, the YOLOv12-large model is employed to detect diseased regions, which obtained an average precision score of 88.5%. These contributions advance the development of accurate, efficient and field-adaptable diagnostic systems for tomato leaf disease management in precision agriculture.
  • Thumbnail Image
    PublicationOpen Access
    Automated design of reinforced concrete dapped-end connections using hybrid deep learning and generative AI augmentation
    (Elsevier Ltd, 2026-04-15) Dharmawansha, S; Herath, S; Fernando P.L.N; Meddage D.P.P.; Rajapakse, C
    Dapped-end connections, also known as half-joints or Gerber beams, are widely used yet structurally vulnerable elements in precast concrete structures due to high stress concentrations near the re-entrant corner. Therefore, a comprehensive assessment of the load-bearing capacity of dapped-end connections is important to ensure structural integrity and mitigate the risk of failure. Although prior studies have explored their behaviour through analytical and experimental methods, the application of data-driven approaches remains limited due to the availability of limited experimental data, which constrains the predictive accuracy and generalisation of Machine Learning (ML) models. This study presents a novel approach that integrates numerical simulation with Conditional Tabular Generative Adversarial Network (CTGAN)-based data augmentation to enhance prediction accuracy and model generalisation. A numerical database containing 720 results was developed, which was expanded with 680 augmented data using CTGAN. The combined dataset of 1400 instances was used to train Artificial Neural Network (ANN), Genetic Algorithm-ANN (GA-ANN), and Particle Swarm Optimisation-ANN (PSO-ANN) models. The hybrid models outperformed the standalone ANN, with GA-ANN achieving the highest accuracy (testing R2 = 0.961). The trained models were separately validated using 64 unseen experimental datasets, which shows the improved generalisation of the models through augmentation. Shapley Additive Explanations analysis reveals that the GA-ANN model predictions aligned with the principles underlying the compatibility of deformations of dapped-ends. Further, a novel ML-assisted design model was developed, which predicts multiple solutions for a given design problem, assisting in the optimisation of connection design.
  • Thumbnail Image
    PublicationOpen Access
    QPred: A Lightweight Deep Learning-Based Web Pipeline for Accessible and Scalable Streamflow Forecasting
    (Tech Science Press, 2026) Makumbura, R.K; Wijesundara, H; Sajindra, H; Rathnayake, U; Kumar, V; Duraibabu, D; Sen, S
    Accurate streamflow prediction is essential for flood warning, reservoir operation, irrigation scheduling, hydropower planning, and sustainable water management, yet remains challenging due to the complexity of hydrological processes. Although data-driven models often outperform conventional physics-based hydrological modelling approaches, their real-world deployment is limited by cost, infrastructure demands, and the interdisciplinary expertise required. To bridge this gap, this study developed QPred, a regional, lightweight, cost-effective, web-delivered application for daily streamflow forecasting. The study executed an end-to-end workflow, from field data acquisition to accessible web-based deployment for on-demand forecasting. High-resolution rainfall data were recorded with tipping-bucket gauges and loggers, while river water depth in the Aglar and Paligaad watersheds was converted to discharge using site-specific rating curves, resulting in a daily dataset of precipitation, river water level and discharge. Four DL architectures were trained, including vanilla Long Short-Term Memory (LSTM), stacked LSTM, bidirectional LSTM, and Gated Recurrent Unit (GRU), and evaluated using Nash-Sutcliffe Efficiency (NSE), Coefficient of Determination (R2), Root-Mean-Square-Error-Standard-Deviation Ratio (RSR), and Percentage Bias (PBIAS) metrics. Performance was watershed-specific, as the vanilla LSTM demonstrated the best generalisation for the Aglar watershed (R2 = 0.88, NSE = 0.82, RMSE = 0.12 during validation), while the GRU achieved the highest validation accuracy in Paligaad (R2 = 0.88, NSE = 0.88, RMSE = 0.49). All models achieved satisfactory to excellent performance during calibration (R2 > 0.91, NSE > 0.91 for both watersheds), demonstrating strong capability to capture streamflow dynamics. The highest performing models were selected and embedded into the QPred application. QPred was developed as a lightweight web pipeline, utilising Google Colab as the primary execution environment, Flask as the backend inference framework, Google Drive for artefact storage, and Ngrok for secure HTTPS tunnelling. A user-friendly front end utilises range sliders (bounded by observed minima and maxima) to gather inputs and provides discharge data along with metadata, thereby enhancing transparency. This work demonstrates that accurate, context-aware deep learning models can be delivered through low-cost, web-based platforms, providing a reproducible and scalable pipeline for hydrological applications in other watersheds and for practitioners. Copyright