Advancing hyperspectral image analysis: harnessing machine learning for precision vegetation identification in urban areas

Exploiting the advantages inherent in machine learning methodologies, particularly supervised learning models tailored for conventional image classification, this study assesses their efficacy in delineating vegetation in hyperspectral images. The investigation adheres to a dataset comprising 400 spectral signatures, evenly divided between vegetation and non-vegetation categories, each characterized by 380 spectral bands. Five distinct models—KNN, decision trees, support vector machines, random forests, and logistic regression—are scrutinized for their performance, leveraging metrics derived from the confusion matrix and cross-validation. The research adopts a modified version of the CRISP-DM methodology, segmented into four phases: understanding the data and the domain, data preparation, modeling and evaluation, and model deployment. Throughout these phases, various open-source libraries such as spectral, scikit-learn, numpy, pandas, and matplotlib are employed. Results indicate that all five models achieve cross-validation accuracies surpassing 95% in vegetation pixel detection within hyperspectral images, with the KNN model exhibiting superior performance at 99.3% accuracy. Subsequently, the model with optimal performance is deployed on a hyperspectral image encompassing the Manga neighborhood in Cartagena, Colombia, comprising 2250000 pixels and 380 frequency bands, yielding highly effective vegetation pixel detection. This article introduces an approach intended to serve as a benchmark for the identification of diverse materials in hyperspectral images at both academic and industrial levels, utilizing open-source technologies.