Alfaro Mejia, Estefania

Profile Picture

Filters

20241
Reset filters

Settings

Citations

Publication Search Results

Now showing 1 - 1 of 1
  • Thumbnail Image
    Publication
    An artificial intelligence based digital image processing framework for hyperspectral image label assignment from unmixing
    (2024-04-21) Alfaro Mejia, Estefania; Manian, Vidya; College of Engineering; Lugo Beuchamp, Wilfredo E.; Andrade Rengifo, Fabio; Luna Hernandez, Adriana; Department of Electrical and Computer Engineering; Veglia, Alex
    Remote sensing involves the analysis of imagery acquired over Earth collected primarily from high-spectral-content sensors on satellites and aircrafts. This thesis presents novel deep learning algorithms for extraction of material composition in acquired scenes. A novel deep learning algorithm for endmember and fractional abundances maps extraction is presented. The proposed convolutional autoencoder significantly improves upon baseline algorithms, demonstrating superior performance on benchmark datasets such as Samson. The metric used to measure the performance achieved by the endmembers extraction is the Spectral Angle Distance, while for the abundance map estimation, is the Root Mean Square Error. Notably, it excels in endmembers extraction for water (0.060) and soil (0.025), as well as abundance estimation with root mean square error metrics of (0.091) for water and (0.187) for soil compared to ground truth. The developed model exhibits the capability to identify chlorophyll-a in waterbodies, serving as a crucial indicator of varying concentrations of macrophytes and cyanobacteria. Additionally, an ensemble algorithm leveraging textural information enhances spectral information for label assignment, achieving high accuracies across various images, including Jasper (93.73%), and cyanobacteria classification (99.92%). The work also proposes a framework addressing label assignment and improving abundance maps through an ensemble deep learning algorithm. Evaluated on the benchmark dataset Samson, the proposed model consistently outperforms baseline algorithms, the algorithm excels in three abundance maps: water, tree, and soil with values of (0.081), (0.158), and (0.182), respectively, showcasing its robust performance.