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

Area of Science:

• Remote Sensing
• Machine Learning
• Environmental Monitoring

Background:

• Hyperspectral remote sensing offers rich data for land and water monitoring.
• Acquiring groundtruth data for hyperspectral images is time-consuming and resource-intensive.
• Automated methods are needed to overcome groundtruth data limitations.

Purpose of the Study:

• To present a semi-supervised method for labeling and classifying hyperspectral images.
• To reduce the reliance on manual groundtruth data collection.
• To enhance the accuracy and efficiency of hyperspectral image analysis.

Main Methods:

• Unsupervised stage: Image enhancement via feature extraction and clustering for groundtruth generation.
• Supervised stage: Preprocessing (normalization, PCA, feature extraction) followed by an ensemble of machine learning models (SVM, gradient boosting, Gaussian, perceptron).
• Ensemble approach: Majority voting to combine model outputs for final classification.

Main Results:

• Gradient boosting showed the best performance in supervised classification.
• High accuracies achieved: 93.74% (Jasper), 100% (HSI2 Lake Erie), 99.92% (cyanobacteria/algal blooms).
• The method effectively differentiates between blue-green algae and surface scum.
• Cloud-based implementation runs 24x faster than a workstation for Lake Erie images.

Conclusions:

• The presented ensemble method effectively generates labeled data for hyperspectral images lacking groundtruth.
• It offers a robust and efficient solution for accurate land cover and water body classification.
• The approach is particularly valuable for monitoring environmental conditions like harmful algal blooms.