Authors - Manohar R, N Abhishek, Nagesh S, Sumith R, C Balarengadurai Abstract - Water quality monitoring is essential for public health and environmental stewardship. Conventional methods, while effective, are often costly, time-intensive, and require specialized skills. In response to these limitations, this paper explores machine learning as a rapid, scalable solution to classify water quality using key parameters, including pH, turbidity, organic carbon, and contaminants. By implementing algorithms such as Random Forest, SVM, and other advanced models, we seek to enhance the precision of water purity assessments. This paper shows the potential of ML applications in real-time monitoring, addressing the need for accessible, cost-efficient, and accurate water quality solutions suitable for broad deployment across diverse environments.
Authors - Hrudai Aditya Dharmala, Ajay Kumar Thallada, Kovvur Ram Mohan Rao Abstract - Recent advances in vision-language models have demonstrated remarkable multimodal generation capabilities. However, their typical reliance on training large models on massive datasets poses challenges in terms of data and computational resources. Drawing inspiration from the expert-based architecture of Prismer, we propose a novel framework for contextual visual question answering specifically tailored to remote sensing imagery. Our methodology extends the Prismer architecture through a two-stage approach: first, by incorporating a domain-specific segmentation expert trained on remote sensing datasets, and second, by integrating a fine-tuned Large Language Model (Mistral 7B) optimized using Parameter-Efficient Fine-Tuning (PEFT) with QLoRA for remote sensing terminology, with hyperparameters optimized with help of Unsloth framework. The segmentation expert performs the analysis of remote sensing imagery, At the same time, the language model acts as a reasoning expert, combining domain-specific knowledge with natural language understanding to process visual contexts and generate accurate responses. In our framework, the use of the Unsloth fine-tuning approach for the language model helps maintain high performance within the defined scope of remote sensing classes and terminology while avoiding hallucination or deviation from established classification schemas. This opens an exciting direction for making the use of Earth observation data more accessible to end-users, demonstrating significant improvements in accuracy and reliability compared to traditional approaches. Experimental results validate that this architecture effectively balances domain expertise with computational efficiency, providing a practical solution for remote sensing visual question answering that requires substantially fewer computational resources compared to end-to-end training of massive models.
Authors - Jenat Arshad, Afruja Akter, Tanjina Akter, Kingkar Prosad Ghosh, Anupam Singha Abstract - Cryptocurrencies have emerged as a significant financial asset class, attracting global attention for their potential to disrupt traditional financial systems. Due to its extreme price volatility and ability to be traded without the assistance of a third party, cryptocurrencies have gained popularity among a wide range of individuals. This paper presents a comprehensive study of machine learning techniques, particularly deep learning models such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Artificial Neural Network (ANN) in predicting cryptocurrency prices. Furthermore, this study addresses the security and privacy challenges inherent to blockchain technology, upon which cryptocurrencies operate. We predict the prices of popular cryptocurrencies like Bitcoin, and Ethereum, and lesser-known ones like Binancecoin, Litecoin, and Ripple through a hybrid deep learning model. This paper also compares cryptocurrency price prediction with machine learning models like GRU, ANN, and our proposed model Hybrid LSTM-GRU. The results demonstrate the efficacy of machine learning in price prediction, highlighting blockchain's potential to enhance security and privacy in financial transactions. Our model gives the value of MSE, RMSE, MAE and MAPE to determine the forecasting. We’ve also added the manual calculation for each metric and compared the actual price with the predicted price that our model gave.
Authors - Harsha S Khurana, Payal D Joshi Abstract - Stress is a major health concern that significantly affects mental stability and can have adverse effects on physical well-being if prolonged. Early detection of stress can help and prevent health-related issues. Individual stress patterns are detected using a variety of bio-signals, including thermal, electrical, auditory, and visual cues which are invasive methods. But according to the well-known saying statement, "Face is a mirror of mind," one can observe one’s emotion or mental state on one’s face. Based on this, Investigated the potential of using facial expressions as a non-invasive method to detect stress levels. Facial expressions could be analyzed and classified as stress and non-stress by examining facial expressions. To solve this problem we have used pretrained network models - Inception, Xception, MobileNetv2, Vgg19, EfficientNet deep learning models, and Affectnet Dataset for stress detection and also represent the comparative study of networks based on confusion and performance metrics. Testing on a separate set of data of images indicates that the MobileNetv2 and Xception models give more accuracy for stress detection.
Authors - Rinkesh N Parmar, Payal D Joshi Abstract - Cotton, an essential crop for the textile industry and millions of farmers, is vulnerable to diseases that can significantly affect yields and profitability. Traditional methods of disease detection, relying on expert visual inspections, are labour-intensive, time-consuming, and prone to errors, often causing delays in addressing problems. This study investigates the use of Convolutional Neural Networks (CNNs) for automated, early, and accurate detection of cotton diseases. CNNs are effective at extracting hierarchical features from raw image data, making them ideal for image classification tasks. In this approach, a labelled dataset of cotton plant images is utilized to train the CNN model, incorporating data augmentation to enhance variability and generalization. The model employs convolutional layers for feature extraction, max-pooling layers for dimensionality reduction, dropout layers for regularization, and fully connected layers for classification. The Adam optimizer, known for faster convergence, is used during training, along with categorical cross-entropy loss. The evaluation is based on accuracy, precision, recall, and F1-score. The model showed significant improvements in performance. The baseline CNN achieved 92.34% accuracy, but advanced architectures like Hybrid CNN-LSTM, DenseNet-121, ResNet-50, and InceptionV3 enhanced accuracy by 2-3%, along with increased precision, recall, and F1-score. The Hybrid CNN-LSTM model outperformed others, achieving 94.5% accuracy, 93.5% precision, 93.2% recall, and 93.3% F1-score. These results suggest that CNN-based models, particularly Hybrid CNN-LSTM, offer substantial improvements in cotton disease detection. The incorporation of data augmentation and dropout regularization strengthens the model, making it effective for real-time agricultural disease management. Future work will focus on expanding the dataset, improving the model, and implementing it in real-world cotton farming practices.
Authors - Joven A. Tolentino Abstract - The growing demand for electricity necessitates effective monitoring and forecasting of consumption trends. This study employs ARIMA modeling, using data from the Department of Energy, Philippines, to analyze and predict electricity consumption. The forecast for the next two years indicated an 18.99% increase in consumption between 2016 and 2017.To enhance analysis, the predicted data was clustered using the K-Means algorithm to group months with similar consumption patterns. This approach identified periods of high, medium, and low electricity usage, providing valuable insights into peak demand months. Such data-driven findings can guide electricity providers in prioritizing resources and implementing strategies to address fluctuations in consumer demand effectively. This study emphasizes the importance of forecasting and clustering as tools for decision-making to mitigate challenges arising from increasing electricity demand.
Authors - Manmeet Borkar, Suneeta Raykar Abstract - Monitoring biomarkers is essential for patients to effectively manage their health profiles and prevent potential complications. Regular tracking of these indicators allows for timely interventions and better control over one’s health, particularly when the methods employed are non-invasive and grant convenience and comfort to the patient. Conventionally, this monitoring is accomplished in pathology laboratories, by collecting blood samples or finger-pricking, which can be distressing and impractical for regular use. Given the growing need for more accessible and affordable healthcare solutions, the development of a cost-effective non-invasive method has become crucial. We propose the use of machine learning models to enable non-invasive measurement of biomarkers such as Total Cholesterol, Uric acid and Blood Sugar. Several Machine learning algorithms, including Linear Regression, K-Nearest Neighbors (KNN), Decision Tree, Random Forest and Support Vector Regression (SVR), were applied to the datasets constructed using the MAX30102 sensor. The metrics used to evaluate regression models were Mean Square Error (MSE) and Coefficient of determination (R²). The final prediction model was built using the algorithm that yielded the highest Coefficient of determination (R²). A user-friendly interface was developed using Tkinter, allowing the input of sensor values from the MAX30102 sensor. The prediction of biomarker values promotes health awareness and timely alerts against potential complications. The results obtained using this approach were validated against laboratory blood reports, revealing an average offset of less than 10% in the predictions.
Authors - Kirthika. P, M. Suresh, S. Kanagaraj Abstract - This paper explores the social barriers faced by Indian plantation communities. It focuses on how these obstacles impact their well-being, productivity, and social mobility. By analyzing historical, socioeconomic, and cultural factors, the study uncovers the multifaceted challenges plantation workers encounter, including income, education, social position in the community, social networks, migration, exploitation, and working and living conditions. The DEMATEL approach identifies the barriers and analyzes the interrelationships among those that impact social barriers among plantation workers. This paper identified seven barriers of impact from a literature review followed by interviews with experts to interpret the interconnection of barriers and investigate the interrelationships. The result says that income and education are the key barriers impacting the lives of plantation workers in their society. The present study incorporates the DEMATEL approach model to analyze the critical barriers in mapping the social barriers of plantation workers. The DEMATEL approach model is the first attempt to study the interrelationship among the barriers. The research overviews the prevailing issues through field surveys, interviews, and literature reviews. The paper will conclude with actionable recommendations aimed at policymakers, community leaders, and stakeholders to mitigate these barriers and promote a more inclusive and equitable environment for plantation workers.
Authors - Aishani Das, Sobitha Ahila, Sreyashi Dey Abstract - Sentiment analysis within the food industry offers essential insights into customer satisfaction, product perception, and emerging concerns. A novel sentiment classification model is developed for Amazon food reviews, leveraging Sentiments are categorized as positive, neutral, or negative using techniques from Natural Language Processing and Machine Learning. Traditional ML algorithms, such as Logistic Regression, Naive Bayes, and Support Vector Machines, are combined with the BERT deep learning model to enhance classification accuracy. With a dataset of over 500,000 reviews sourced from Kaggle, the methodology includes data cleaning, feature extraction, exploratory data analysis, model training, and evaluation. Initial findings demonstrate SVM’s high predictive accuracy in sentiment classification, while BERT’s advanced contextual understanding suggests further enhancements. Applications of this model extend to real-time feedback systems that assist businesses in identifying and addressing customer sentiments promptly. Future developments aim to improve accuracy, incorporate a diverse range of datasets, and integrate real-time processing and multilingual analysis for broader, more effective sentiment analysis capabilities.
Authors - Tinoy Santra, Sahil Neekhra, Ritik Gupta, Gunabalan Ramachandiran Abstract - With the rising environmental degradation and increasing global warming, electric vehicles are the promising concept in the automobile industry. Different sources of energy are available for giving power to drive the vehicle. Sunlight being an efficient and abundant resource, the world is moving towards solar energy leaving behind conventional power resources. Moreover, battery based electric vehicles have short driving range and speed which is not acceptable in the dog-eat-dog market. This paper discusses a simple approach for BLDC motor driven electric vehicle powered by buck-boost converter. The primary energy source is solar energy, and the PI controller holds the DC-DC converter's output constant. A 660 W, 48 V BLDC motor driven electric bike system is worth an alternative when it is solar powered which solves utmost all the problems faced in usage of EVs. The circuit is simulated in MATLAB environment and output parameters are observed for different load conditions. Overall, the motive is to prove that electric vehicles are more efficient and cost effective than the conventional ones.
