Unlike classical color images with 3 bands, Hyperspectral images (HSIs) can provide detailed spectral information about the imaged objects in hundreds of narrow bands, allowing this way differentation between materials that are often visually indistinguishable. Classification of HSIs gains currently lots of attention in remote sensing community. The objective of supervised hyperspectral classification is to group pixels into different classes (such as buildings, roads, parking lots, etc) with the classifiers trained by the given training samples.
We focus on the classifiers based on sparse representation (SR) technique, which has been successfully applied to numerous applications, especially in the fields of computer vision, image processing, pattern recognition and machine learning. Sparse representation classification (SRC) was first applied in face recognition and achieved a significant success over more traditional classifiers, such as support vector machines (SVM). The main underlying idea is that each test sample can be well represented as a linear combination of relatively few atoms from a large dictionary. SRC techniques construct these dictionaries of signal atoms directly from the training samples. Then by calculating the class-specific reconstruction errors, the class of each test sample can be identified.
A number of SRC-based methods have been proposed in the last several years. However, these existing methods construct a dictionary by randomly selecting training samples regardless of their discrimination power, which limits the performance in practice. It is therefore important to develop new and smarter schemes for sampling representative atoms from the training data and to improve thereby the discrimination between different classes of input samples. In addition, real data are typically affected by various types of noise, which hamper severely the classification performance, and this aspect is also not sufficiently well addressed in the current literature. Hence, various aspects of the emerging SRC-based methods are yet to be explored, including effective dictionary design, robust classification model building and efficient implementation of the algorithms. The research group IPI is active in this domain and will provide full support in terms of the available software, reference techniques and case studies for validation.
The goal of this Master thesis is to improve further one of the currently most popular and emerging approaches in hyperspectral image classification, based on sparse representation techniques. The research will mainly address the following issues:
The students will be able to start from the current software for the SRC-based classification developed in the research group GAIM and will also obtain the training and validation data from real applications in remote sensing. Motivated students will be encouraged to participate with the developed techniques in data classification challenges of the IEEE Geoscience and Remote Sensing community.