Sylhet Engineering College

Traditional metasurfaces often face challenges in achieving broadband functionality and dynamic adaptability, limiting their use in advanced electromagnetic systems. This paper presents a triple circular ring metasurface designed for multifunctional electromagnetic applications, including surface wave propagation, beam shaping, and ultra-broadband radar cross-section (RCS) reduction. The proposed structure uses a cost-effective FR-4 substrate and demonstrates strong electromagnetic reflection characteristics across 50-100 GHz. Except near 91 GHz, the metasurface exhibits amplitude and phase responses comparable to a conventional copper plate while maintaining efficient surface wave propagation. Significant electric and magnetic field amplitudes of nearly 1 V/m and 5x10^-3 A/m are sustained across the surface, unlike a standard copper plate. The metasurface also redirects incident energy toward a predefined direction of 67 degrees in the phi plane while minimizing radiation over a 360-degree angular range. In addition, it achieves a stable monostatic RCS reduction from -40 dB to -30 dB across a broad frequency range, outperforming conventional copper structures. Numerical simulations validate the proposed design. The results demonstrate strong potential for stealth technology, radar systems, and next-generation wireless communications.

The encroachment of vegetation along power lines is a major cause of grid outages and even safety risks in power line corridors. Traditional surveillance systems are expensive and can be extremely infrequent that require automated and highquality analysis. In this paper, we present a single deep learning backbone to perform semantic segmentation of vegetation, power lines and background of Unmanned Aerial Vehicle images. To train an effective end-to-end model to overcome the severe class imbalance intrinsic to this task, especially the problem of thin power lines segmentation, we use UNet architecture with a pretrained ResNet34 backbone. With a composite loss form based on weighted Cross-Entropy and multiclass Dice loss, our singlemodel system provides a much simpler workflow than the later multi-model ensemble models. Our structure shows the state of the art performance with respect to publicly available benchmark dataset as evaluated using Intersection over Union, Precision and Recall measures. The model achieves the overall mean IoU of 0.9877 on real-world test data, which is further supported with Precision and Recall values across all classes, with the most notable values associated with the power line class. It is an accurate, efficient and practical approach to automated monitoring of the power line infrastructure which is a valuable tool in improving the resilience of the grid and avoiding failures that happen due to vegetation.

Timely and accurate identification of poultry diseases is crucial for maintaining livestock health and ensuring sustainable productivity. This study proposes PoultryXNet, a novel lightweight Convolutional Neural Network (CNN) framework for automated poultry disease classification from fecal images. PoultryXNet integrates complementary features from MobileNetV3-Small, ShuffleNetV2, and EfficientNet-B0, enhancing discriminative capability while maintaining computational efficiency. The model was trained and evaluated on a dataset comprising $\mathbf{6, 8 1 2}$ poultry fecal images across four categories (Coccidiosis, Newcastle Disease, Salmonella, and Healthy), encompassing both diseased and non-diseased samples. To mitigate class imbalance, extensive data augmentation was applied across all classes. PoultryXNet achieved a validation accuracy of 99.88% and a test accuracy of 99.62%, with corresponding precision, recall, F1-score, and AUC-ROC of $99.63 \%, 99.63 \%, 99.62 \%$, and 0.999984, respectively. These results outperform several state-of-the-art baseline models, including VGG16, ResNet50, EfficientNet-B0, and MobileNetV2. Furthermore, Grad-CAM++ visualizations highlight disease-specific regions in fecal images, providing interpretability and enhancing trust in the proposed model's decisions. This framework provides a highly accurate, efficient, and explainable solution for diagnosing real-world poultry diseases.

The Buriganga River, the lifeline of Dhaka, has suffered severe pollution due to rapid urbanization and industrial discharge. This study evaluates surface-water quality and pollution dynamics from 2010 to 2023 using a GIS-based multi-criteria approach. Three Water Quality Index (WQI) methods such as National Sanitation Foundation (NSF), Canadian Council of Ministers of the Environment (CCME), and the Weighted Arithmetic (WA) were applied to long-term monitoring data from four key stations (Mirpur Bridge, Hazaribagh, Kamrangir Char, Sadarghat). In addition, Comprehensive Pollution Index (CPI) and Organic Pollution Index (OPI), were also computed to capture overall and organic pollution loads. The findings present consistently poor water quality, with NSF-WQI values of 22.35–31.16 (“Very Bad – Bad”), CCME-WQI 41.60–52.38 (“Marginal – Bad”), and WA-WQI 310.41–606.13 (non-potable). Both CPI and OPI values indicate severe contamination, particularly during the dry season, with Hazaribagh and Kamrangir Char being the most polluted zones. Spatial–temporal mapping confirms industrial and domestic effluents as dominant stressors. Among the evaluated models, the CCME-WQI proved most suitable for the Buriganga context due to its sensitivity to seasonal and spatial variability. The findings highlight an urgent need for strengthened effluent regulation, sustainable wastewater management (SDG 6), and continuous GIS-based monitoring to reestablish this vital urban river (SDG 11).

This research investigates the application of deep convolutional neural networks (CNNs) for classifying 5 prominent rice varieties in Bangladesh: Basmati, Chinigura, Jirashail, Kataribhog, and Paijam. Rice is a significant agricultural product in Bangladesh. Proper categorization of rice varieties is vital for quality management, preventing adulteration, and preserving customer confidence. This research leverages advanced CNN architectures (VGG16, VGG19, ResNet50, Xception, DenseNet) and a custom model, BDriceNetworkV0, to automate the classification process. A balanced dataset comprising 15 000 images of rice grains was collected under controlled conditions to train the models. Performance metrics, including accuracy, precision, recall, and F1-score, were used to evaluate the models. VGG19 was identified as the superior performer (accuracy 99.73