General Attractions for Hyperspectral Data - cpshooter/geoML GitHub Wiki

The analysis of HSI turns out to be more difficult, especially because of the high dimensionality of the pixels, the particular noise and uncertainty sources observed, the high spectral redundancy, and the typically nonlinear relations observed between spectral channels as well as with the corresponding material. Such nonlinearities can be related to a plethora of factors, including the multiscattering in the acquisition process, the heterogeneities at subpixel level, as well as the impact of atmospheric and geometric distortions. These characteristics of the imaging process lead to distinct nonlinear feature relations, i.e., pixels lie in high-dimensional complex manifolds. The high spectral sampling of HSI (the bands usually cover narrow portions of the electromagnetic spectrum, typically 5–10 nm) also leads to strong collinearity issues. Finally, the spatial variability of the spectral signature increases the internal class variability.

All of these factors, in conjunction to the few labeled examples typically available, make HSI image classification a very challenging problem.

HYPERSPECTRAL imaging sensors are capable of capturing fine and abundant spectral information in hundredths of continuous narrow spectral bands. The interests for image processing techniques and wide range of applications of hyperspectral images have been greatly increased over the last decades. One of the most important processing tasks is classification, by which land-cover thematic maps are generated [1]. However, there are several important challenges when performing hyperspectral image classification. For instances, unbalance between the limited training samples and high dimensionality, presence of mixed pixels, integrating the spatial and spectral information to take advantage of the complementarities, quite complex geometry, and very high computational complexity of some of the classifiers [2]. To overcome these challenges and perform hyperspectral image classification effectively, several machine learning techniques such as artificial neural networks (ANNs) [3], support vector machine (SVM) [4]–[8], multinomial logistic regression [9], active learning (AL) [10, 11], semisupervised learning (SSL) [12], manifold learning [13], etc., and other methods like hyperspectral unmixing via sparse representation [14], morphological profiles [15], [16], and partitional clustering techniques [17] have been popularly investigated in recent years as well. However, due to the high dimensionality and complexity of hyperspectral image in spectral domain, finding optimal parameters for parametric supervised algorithms is always time-consuming and difficult. It is still a critical problem to achieve high-performance classification with fast speed and high efficiency.