Researches

Undergraduate Thesis

IoT-based smart water quality monitoring and automatic control system designed for urban pipeline networks. The system continuously measures four critical water quality parameters—pH, temperature, turbidity, and Total Dissolved Solids (TDS)—using sensors connected to an Arduino-based setup. Sensor data is sent to a Firebase real-time cloud database, visualized on a custom web dashboard, and monitored remotely by users and authorities.

If any water parameter exceeds WHO-recommended safety limits, the system instantly:

• Stops water supply by closing a solenoid valve

• Sends alert SMS through a GSM module

• Displays real-time warnings on the website

To evaluate performance, the authors collected 10,000+ sensor data samples, labeled them as safe/unsafe, and trained multiple machine learning models. A Decision Tree classifier achieved 99% accuracy, showing strong potential for predictive, automated water safety assessment.

The study concludes that this low-cost smart system can significantly reduce waterborne diseases and improve water supply management in smart cities. Future improvements include mobile app integration, leakage detection, in-pipeline purification, and disease prediction from sensor data.

Transformer-based models for detecting abusive Tamil and Malayalam text targeting women on social media, a challenging task for low-resource Dravidian languages. The work evaluates several multilingual transformers—XLM-R, MuRIL, IndicBERT, mBERT, and ensemble methods—and identifies the most effective models for each language.

Using the official DravidianLangTech@NAACL 2025 dataset, the study finds that:

• XLM-RoBERTa performs best for Tamil, achieving an F1 score of 0.7873, ranking 2nd in the shared task.

• MuRIL performs best for Malayalam, achieving an F1 score of 0.6812, ranking 10th.

The results highlight that multilingual transformers can handle abusive text in low-resource languages, but challenges remain in handling sarcasm, dialect variations, and contextual ambiguity. The paper contributes an effective approach for gender-targeted abuse detection and offers insights for improving NLP models in Dravidian languages.

This work introduces a high-quality dataset of 6,419 images of six traditional medicinal plant species collected from the natural jungle habitat of Guadanga, Phulpur, Mymensingh, Bangladesh. The dataset includes original, annotated, and augmented leaf images, captured using Google Pixel 5 and 6a smartphones under controlled lighting. All images were preprocessed to 640×640 resolution and manually annotated with bounding boxes in YOLO format.

The dataset features six widely used medicinal plants Centella asiatica, Coccinia grandis, Eclipta prostrata, Mikania micrantha, Murraya koenigii, and Stephania japonica providing realistic morphological diversity rarely found in existing datasets. To validate quality, a YOLOv11 model was trained on the dataset and achieved an outstanding mAP@50 of 99%, confirming its suitability for leaf classification and object detection tasks.

Publicly available on Mendeley Data, this dataset serves as a valuable benchmark for computer vision, deep learning, ethnobotany, agricultural informatics, and AI-based medicinal plant identification, supporting future research in digital herbariums and biodiversity documentation.