Orozco Sandoval, Jairo J.

Profile Picture

Filters

2019 - 201912020 - 20251
Reset filters

Settings

Citations

Publication Search Results

Now showing 1 - 2 of 2
  • Thumbnail Image
    PublicationRestricted
    Discovering phases and phase transitions using machine learning
    (2019-05-14) Orozco Sandoval, Jairo J.; Ramos Maldonado, Rafael A.; College of Arts and Sciences - Sciences; Jiménez González, Héctor J.; Lu, Kejie; Li, Yang; Department of Physics; Alers Valentín, Hilton
    Machine learning a specific subset of artificial intelligence, trains a machine to learn from data. It has become a robust method for the identification of patterns within complex physical systems to determine certain physical quantities without prior knowledge of their physics principles. In this thesis work, we apply an unsupervised machine learning technique, Principal Component Analysis, and a supervised learning technique, Artificial Neural Networks, to identify phases and phase transitions in square and hexagonal lattice Ising models. It can be drawn from the results that Principal Component Analysis can successfully identify phase transitions and locate the transition temperatures in both square and hexagonal lattice systems. Additionally, it was found that two principal components are related to the order parameter and the susceptibility of the systems. The weight vectors have, then, a physical explanation, which is helpful to better understand system behavior. On the other hand, by employing neural networks, it was possible to understand the training effects of the Ising model, as well as obtain a critical temperature value ??C close to the real thermodynamic value. This confirms machine learning is suitable for approaching the type of complex systems studied in this research. Principal Component Analysis and neural network can learn from complex data, without the need of significant human intervention.
  • Thumbnail Image
    PublicationEmbargo
    Hyperspectral image representation and classification using graphs and graph convolutional neural networks
    (2025-04-12) Orozco Sandoval, Jairo J.; Manian, Vidya; College of Engineering; Ramos Maldonado, Rafael A.; O’Neill Carrillo, Efraín; Vega Riveros, José F.; Department of Electrical and Computer Engineering; Hernández Hernández, Carlos I.
    Hyperspectral image (HSI) classification is a challenging task due to the high di- mensionality, spectral redundancy, and complex spatial structures present in the data. Traditional deep learning models, such as Convolutional Neural Networks (CNNs), often fail to capture the non-Euclidean relationships among pixels, leading to suboptimal classification performance. Graph Convolutional Networks (GCNs) offer a promising alternative by modeling HSIs as graphs, where nodes represent pixels or superpixels, and edges encode spectral-spatial relationships. This thesis investigates the effectiveness of GCNs for HSI classification by constructing and optimizing graph representations. Several similarity metrics, including Spectral Information Divergence (SID), Spectral Angle Mapper (SAM), Wasserstein Distance (WD), and Radial Basis Function (RBF), are explored to generate adjacency matrices that accurately capture pixel-wise depen- dencies. Additionally, an adaptive neighborhood selection mechanism is developed to dynamically refine graph connectivity and improve classification accuracy. To evaluate the proposed methodologies, several experiments are conducted on multiple benchmark datasets, including Indian Pines, Pavia University, Salinas, Botswana, and University of Houston. An ablation study is performed to assess the impact of different adjacency matrix constructions and neighborhood selection strategies. Results demonstrate that GCNs leveraging adaptive graph representations significantly outperform traditional CNNs and other GCN based deep learning methods, particularly in scenarios with lim- ited training samples. Furthermore, a novel dimensionality reduction approach based on Local Linear Embeddings (LLE) is integrated with GCNs to enhance computational efficiency without compromising classification accuracy. Hybrid architectures, incorporating autoencoders based in deep neural networks (DNNs) and CNN with GCN for accuracy improvement are presented. Among the proposed models, the S3CLLE-WGCN framework achieves state-of-the-art performance across multiple datasets, setting new benchmarks in HSI classification. The findings of this research highlight the advan- tages of leveraging graph-based learning techniques for HSI analysis. The integration of similarity measures for adjacency matrix construction, spectral, and spatial information into graph structures enables a more expressive representation of HSI data, paving the way for improved classification accuracy in practical applications such as precision agriculture, environmental monitoring, and land cover mapping.