University of Puerto Rico at Mayagüez Institutional Repository
Welcome to the University of Puerto Rico at Mayagüez Institutional Repository.
Bienvenidos al Repositorio Institucional de la Universidad de Puerto Rico en Mayagüez.
Featured Items
Recent Submissions
Publication Intrinsic dimensionality kernels based feature learning strategies for hyperspectral images(2025-05-12)This thesis introduces a novel framework for hyperspectral image classification, by preserving the intrinsic spatial-spectral relationships and emphasizing intrinsic spatial-spectral relationships and manifold geometry in high-dimensional data. Unlike traditional deep learning methods that typically assume Euclidean structures, this approach leverages manifold characteristics and intrinsic dimensionality, preserving essential spectral and spatial features during dimensionality reduction. A central innovation is the Supervised Spatially Spectrally Coherent Local Linear Embedding (S\textsuperscript{3}CLLE), which integrates label information and spatial coherence into the embedding process. This method enhances local geometry preservation and class separability, significantly improving classification accuracy. Additionally, a distance-based sparse adjacency matrix for Graph Convolutional Networks (GCNs) is proposed, demonstrating that using Wasserstein distance refines neighborhood definitions and boosts performance, particularly in few-shot learning scenarios. Combining S\textsuperscript{3}CLLE embeddings with this adaptive graph structure yields a substantial improvement in classification reliability. Experimental evaluations on benchmark datasets (Indian Pine, Pavia University, Salinas, Houston University) confirm superior accuracy and computational efficiency over state-of-the-art methods. The results highlight that incorporating manifold geometry, spatial coherence, and advanced adjacency metrics significantly advances hyperspectral image analysis, setting the groundwork for future high-dimensional data processing.Publication Photogrammetry to quantify the structural complexity provisioning by deep-sea corals(2025-05-10)Structural complexity is a key ecosystem function that enhances biodiversity, habitat availability, and ecological interactions. Keystone taxa like deep-sea corals play major roles in generating this complexity, yet quantifying their contributions remains challenging. We developed a novel method to assess the structural contributions of Hawaiian deep-sea corals at the Keāhole Precious Coral Gardens (natural reef) and the I-201 Submarine (artificial reef) by generating 3D models from high-resolution imagery and digitally removing corals to compare structural complexity (defined as the ratio of 3D to 2D surface area) between original and coral-free models. We evaluated site-level coral contributions, taxa-specific complexity, observer bias in volume estimates due to canopy effects, and scale dependence of structural complexity. Corals enhanced structural complexity by 13% at Keāhole and 3% at Submarine. Taxa-specific contributions varied: Kulamanamana haumeaae and Paracalyptrophora hawaiiensis at Keāhole generated higher complexity per unit height than Narella spp. at the Submarine. Canopy-related observer bias in 2D orthomosaics obscured over 50% of coral volume at Keāhole, compared to less than 3% at Submarine, highlighting limitations of 2D analysis in structurally complex habitats. Structural complexity increased exponentially at finer scales, with coral influence detectable at areas smaller than 10 cm². This method provides a non-invasive approach to quantify the structural role of habitat-forming taxa and can be applied across ecosystems to assess how keystone species shape habitat structure and biodiversity.Publication Design, development and analysis of Raman spectroscopy multivariate and machine learning methods for non-invasive, real-time monitoring of cell culture media and supernatants in a cardiac differentiation system(2025-05-09)The complexity of cellular manufacturing processes necessitates advanced monitoring tools to ensure the quality of cell-based products for regenerative medicine. The use of cell culture media is important for manufacturing process because it is a complex mixture essential for cell growth, differentiation, and metabolic processes. This study evaluates the integration of Raman spectroscopy with multivariate and machine learning to monitor cellular media, glucose quantification, and cardiac differentiation from human-induced pluripotent stem cells (hiPSCs). Raman spectra of powdered cell culture media were analyzed to identify chemical variations related to biomolecules. Principal Component Analysis (PCA) differentiated media based on glucose and pyruvate concentrations, while Partial Least Squares (PLS) regression quantified glucose in complex mixtures, mitigating heterogeneity with composite sampling. For cardiac cell differentiation from human-induced pluripotent stem cells (hiPSCs), Raman spectroscopy identified biomolecular variations in cell supernatants collected at different process stages. While PCA highlighted trends related to differentiation on day 15, machine learning models—Random Forest, Deep Neural Networks, and K-Nearest Neighbors—showed superior performance, achieving over 80% accuracy in classification tasks. These models demonstrated potential for early-stage differentiation monitoring, enabling quality control and cost reduction in cell manufacturing. This work emphasizes the utility of Raman spectroscopy coupled to multivariate analysis and machine learning as a tool for quality control and early prediction of cell potency in a cell manufacturing process offering a low-cost, at line, non-invasive, real-time solution for improving the consistency and reliability of cellular manufacturing processes.Publication Rapid, portable and low-cost water quality device using machine learning(2025-05-08)Maintaining high water quality standards is essential for preventing the spread of waterborne diseases that pose significant risks to public health, such as cholera, dysentery, hepatitis, and typhoid fever. While conventional methods, including membrane filtration (MF), multiple tube fermentation (MTF), and enzyme-based assays, offer high sensitivity and specificity in detecting bacterial indicators like Escherichia coli (E. coli) and Enterococcus faecalis (E. faecalis), these techniques are hindered by lengthy processing times, reliance on specialized laboratory facilities, and the need for trained personnel. To address these limitations, this thesis presents the development of a novel, portable, and cost-effective UV-LED/RGB sensor system for rapid bacteriological water quality assessment. The system incorporates a multi-well, self-loading microfluidic device, UV-LEDs for sample excitation, and RGB sensors (devices capable of detecting light intensity in red, green and blue wavelengths ranges) for fluorescence data acquisition, alongside a defined substrate assay and portable incubation system for automated bacterial quantification. The microfluidic device autonomously loads water samples, eliminating the need for sample preparation, and enables bacterial enumeration via most probable number (MPN) analysis. Fluorescence signals from individual wells are processed using machine learning (ML) algorithms, including Multilayer Perceptron Neural Networks (MLPNN), Random Forest (RF), Logistic Regression (LR), and Support Vector Machines (SVM). These algorithms classify well as positive or negative within 30 minutes. The results indicate that the best-performing algorithm was MLPNN, achieving evaluation metrics of up to 100% for E. faecalis detection. For this study, E. coli was not utilized as an indicator. Comparisons with the Quanti-TrayTM/2000 system revealed greater dispersion and limitations in the detection range for high bacterial concentrations. However, the rapid response time, the reduced sample volume required, and the integration of automation and ML-based analysis tools present a promising alternative for bacteriological detection and quantification.Publication The impact of the COVID-19 pandemic on the teaching experience of English instructors at the University of Puerto Rico at Mayagüez(2025-05-06)Many epidemics like The Spanish Flu, Smallpox, Tuberculosis, among others have impacted Puerto Rico. However, the Coronavirus (COVID-19) shook the entire world leaving serious consequences. For that reason, this study focused on how COVID19 affected the teaching experience of three English professors at a public university in Puerto Rico. This qualitative research utilized life stories to have a better understanding of the participant’s circumstances and outcomes. The findings evidenced that lesson planning, lack of administrative support, lack of clarity boundaries between educators’ professional and personal lives were a source of stress that influenced their teaching. During this time, uncertainty plagued professors, students and the administration alike. These unforeseen events caused institutions to change their policies multiple times to comply with mandatory government requirements. The establishment of emergency policies is vital to ensure the safety of the faculty and students during possible future events of this magnitude to ensure academic success.
Communities in Scholar@UPRM
Select a community to browse its collections.