Hyperspectral Image Labeling and Classification Using an Ensemble Semi-Supervised Machine Learning Approach

Abstract

Hyperspectral remote sensing has tremendous potential for monitoring land cover and water bodies from the rich spatial and spectral information contained in the images. It is a time and resource consuming task to obtain groundtruth data for these images by field sampling. A semi-supervised method for labeling and classification of hyperspectral images is presented. The unsupervised stage consists of image enhancement by feature extraction, followed by clustering for labeling and generating the groundtruth image. The supervised stage for classification consists of a preprocessing stage involving normalization, computation of principal components, and feature extraction. An ensemble of machine learning models takes the extracted features and groundtruth data from the unsupervised stage as input and a decision block then combines the output of the machines to label the image based on majority voting. The ensemble of machine learning methods includes support vector machines, gradient boosting, Gaussian classifier, and linear perceptron. Overall, the gradient boosting method gives the best performance for supervised classification of hyperspectral images. The presented ensemble method is useful for generating labeled data for hyperspectral images that do not have groundtruth information. It gives an overall accuracy of 93.74% for the Jasper hyperspectral image, 100% accuracy for the HSI2 Lake Erie images, and 99.92% for the classification of cyanobacteria or harmful algal blooms and surface scum. The method distinguishes well between blue green algae and surface scum. The full pipeline ensemble method for classifying Lake Erie images in a cloud server runs 24 times faster than a workstation.