Complex networks for modeling texture and spectral features of hyperspectral images for environmental analysis

A complex network method is introduced for high-resolution hyperspectral image analysis and classification. The method is applied to detecting environmental pollution with the Jacaranda caroba plant species. Using confocal laser scanning microscopy (CLSM), detailed spectral data were captured from leaves exposed to different levels of potassium fluoride. Unlike most studies that focus on pixel- or patch-level classification, this work targets the classification of entire high-resolution hyperspectral images, requiring a method capable of capturing global spatial-spectral and texture relationships. Therefore, the limited number of samples and the high dimensionality of the hyperspectral data make conventional deep learning methods unsuitable, motivating the need for a robust and efficient alternative. To address this, we developed the hand-engineered technique named Directed Network of Angular Similarity (DNAS) which models the hyperspectral pixels as complex network vertices connected based on the angular similarity of their spectral bands. This technique allows for effective and efficient feature extraction, computing a compact image representation with only 36 descriptors. Coupled with a supervised classifier, our method achieves a classification accuracy of 92.6% when distinguishing Jacaranda caroba pollutant levels, surpassing both traditional and deep learning approaches. By leveraging the structural, spectral, and texture properties of hyperspectral data, the DNAS method provides a novel framework for detecting pollutant-induced changes in leaf structure, offering significant advantages in resource-limited scenarios. The results demonstrate the potential of Jacaranda caroba leaves, analyzed with this innovative technique, to serve as indicators of air quality.