http://drr.vau.ac.lk/handle/123456789/1299 2026-04-05T19:48:37Z http://drr.vau.ac.lk/handle/123456789/1420 Deep Learning for Intelligent Waste Sorting in Sri Lanka: A Conceptual Review Herath, H.M.P.H.S.; Kuruneru, R.P.D.; Logeesan, R.; Venuja, N. Urban waste management in Sri Lanka is increasingly challenged by rising waste volumes, reliance on manual sorting, and inefficient collection systems. This review paper aims to examine how deep learning, particularly Convolutional Neural Networks (CNNs), has been applied in global studies for intelligent waste classification and to assess its potential applicability in the Sri Lankan context. A systematic survey of recent research highlights models such as lightweight CNNs, VGG16, Dense Net, and hybrid CNN-SVM approaches, which report accuracies exceeding 90% across datasets like Trash Net. Techniques including transfer learning, data augmentation, and optimization methods have been widely adopted to improve model generalization and real-time usability. The review identifies key opportunities for Sri Lanka, including the use of localized datasets, cost-effective deployment without reliance on IoT infrastructure, and mobile or web-based platforms for citizen and municipal use. However, most reported studies are based on controlled datasets not reflective of Sri Lanka’s waste profile, and no significant local implementation has yet been reported. This limitation emphasizes the need for future work on dataset creation, model adaptation, and real-world evaluation under local conditions. The paper concludes that while deep learning offers strong potential to enhance waste classification, further research and pilot projects are essential to translate these global advancements into practical, sustainable solutions for urban Sri Lanka. 2025-01-01T00:00:00Z http://drr.vau.ac.lk/handle/123456789/1419 Sense Step: An Innovative Smart Cane for Visually Impaired Individuals Tharsayani, M.; Thisakaran, R. Sense Step is an innovative smart cane designed to empower visually impaired individuals through advanced assistive technology. Traditional mobility aids lack critical features to ensure user safety and independence, leaving visually impaired people vulnerable to obstacles, health emergencies, and navigation difficulties. Sense Step addresses these challenges by integrating IoT sensors, AI-driven navigation, and real-time health monitoring into an intuitive, user-centric device. The cane utilizes ultrasonic sensors and LiDAR technology for accurate obstacle detection, while embedded health sensors monitor vital signs like heart rate and blood pressure. An accelerometer-based fall detection system automatically alerts emergency contacts if a sudden fall occurs, providing an added layer of security. To ensure uninterrupted functionality, the device features a solar-powered charging system and long-lasting battery life. The smart cane connects to a cloud-based platform via Wi-Fi and Bluetooth, enabling seamless data synchronization and remote monitoring through a dedicated mobile app. Developed using embedded systems (ESP32 microcontroller) and sensor fusion techniques, Sense Step delivers reliable performance in real-world conditions. Its ergonomic, waterproof design ensures durability and comfort for daily use. The result is a cost-effective, scalable solution that enhances mobility, reduces dependency on caregivers, and improves overall quality of life for visually impaired users. By combining cutting-edge technology with practical usability, Sense Step redefines assistive mobility, offering a safer, smarter alternative to traditional canes. With future potential for AI-powered navigation and smart city integration, the device represents a significant step toward inclusive, independent living for the visually impaired community 2025-01-01T00:00:00Z http://drr.vau.ac.lk/handle/123456789/1418 Harvest Guardian: A Mobile Toolkit for AI-Based Crop Disease Diagnosis and Farmer Decision Support Thivakaran, V.; Thisakaran, R. Agriculture in Sri Lanka is predominantly driven by smallholder farmers who face challenges in timely disease detection, access to localized expert advice, and planning around weather conditions. These issues contribute to significant crop losses and reduced household income, highlighting the need for accessible, technology-driven solutions. This paper presents Harvest Guardian, a mobile toolkit designed to support farmers through integrated, user-friendly features. The system incorporates a pretrained TensorFlow Lite model for offline crop disease detection across 30 classes, supported by a multilingual disease wiki with text-to-speech functionality. Additional modules include a peer-based community forum, a localized weather forecasting service, and a marketplace for trading agricultural goods. Firebase services provide secure authentication, media handling, and real-time data synchronization. Internal validation demonstrated that all core modules operated stably across multiple Android devices, with the disease detection component offering reliable offline functionality. While the model has not been retrained with local datasets and field testing remains pending, the architecture establishes a strong foundation for future expansion. Planned improvements include dataset localization, structured usability testing, and wider deployment, positioning Harvest Guardian as a scalable and practical solution for enhancing farmer decision-making in low-resource environments. 2025-01-01T00:00:00Z http://drr.vau.ac.lk/handle/123456789/1417 Detecting Pests in Mango Crops using Machine Learning (ML) Sankeetha, V.; Agalya, V.; Vinoharan, V. Mango is one of the most important fruit crops in Sri Lanka, playing a vital role in both the agricultural sector and rural livelihoods. Despite its economic value, mango cultivation faces persistent challenges due to pest infestations, which significantly reduce yield and fruit quality. Early detection of pests is critical to minimize damage, but manual inspection methods are labor-intensive, time-consuming, and often impractical for largescale farms. With advancements in artificial intelligence, automated detection methods provide a promising alternative to improve pest management efficiency and accuracy. This study introduces a machine learning-based approach for the automated detection of pests on mango leaves, aiming to support farmers and agricultural professionals with a practical tool for timely intervention. The research utilizes ten classes from the Mango Pest Classification dataset, which contains real-world images of healthy and pest-affected mango leaves captured under diverse environmental conditions, thereby reflecting realistic field scenarios. To prepare the data for classification, three feature extraction techniques were applied: Histogram of oriented Gradients (HOG) to capture shape and edge information, Bag of Features (BoF) to model local descriptors, and Wavelet transforms to analyze texture and frequency components. These extracted features were then used to train and evaluate three machine learning classifiers: Support Vector Machines (SVM), Random Forest, and Logistic Regression. The experimental results showed that the Random Forest classifier, when trained on a combined feature set of HOG and Wavelet descriptors, achieved the highest accuracy of 81% on the test data. This performance outperformed other feature-classifier combinations, highlighting the robustness of Random Forest for pest classification tasks. The findings emphasize the potential of machine learning to enhance pest monitoring in mango cultivation. By reducing reliance on manual inspection and enabling timely pest detection, the proposed approach can contribute to improved yield, higher fruit quality, and more sustainable mango farming practices in Sri Lanka. 2025-01-01T00:00:00Z