EVALUATION OF MACHINE LEARNING FOR HYPERSPECTRAL REMOTE SENSING IMAGE CLASSIFICATION

As technology advances, hyperspectral remote sensing image has become an exciting area of research. With the advantages of hyperspectral images (HSI), such as detailed spectral information, there is a great potential for sophisticated applications across diversified fields, including enhancing image classification accuracy. Hyperspectral imagery may achieve more precise object differentiation. However, dealing with high-dimensional hyperspectral data presents challenges, including the curse of dimensionality and redundant information between bands. To overcome these obstacles, integrating machine learning techniques offers effective solutions. In this research, feature extraction is performed using the Minimum Noise Fraction (MNF) technique to obtain relevant and important information. Subsequently, three machine learning classification algorithms, including Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost), are applied for classification. The research utilizes two airborne hyperspectral benchmark datasets, Indian Pines and Salinas, obtained from JPL's Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) project. The datasets consist of samples from various land cover categories, including forests, agricultural areas, bare soil, and vegetation, with samples representing different growth stages. Both datasets contain 16 classes of samples but cover different categories. Additionally, the study intends to incorporate EO-1 Hyperion hyperspectral satellite image, which provides a broader coverage area, serving as supplementary research data. Moreover, SPOT imagery data is utilized as a reference to evaluate the results of Hyperion classification. The objective of this research is to integrate the advantages of various algorithms and find a suitable and efficient classification process for hyperspectral image data, which can be successfully applied to both airborne and satellite images. Experimental results show that this process can achieve satisfactory classification results for both airborne and satellite hyperspectral images. All three classification algorithms produce acceptable results and the Random Forest algorithm produces the best classification accuracy. The best results show an overall accuracy of 0.86 and 0.96 for Indian Pines and Salinas airborne images, respectively. The best overall accuracy of the satellite image is 0.84 and Kappa is 0.79.