A block-compressive sensing algorithm for hyperspectral data processing and acquisition: A comparative classification performance study

Abstract

Current deployments of remote sensing hyperspectral imaging (HSI) technology usually require expensive setup costs, specialized aircraft and associated infrastructure, in addition to producing amounts of data that have a nontrivial impact on data storage resources. Compressive Sensing (CS) and sparse recovery of signals is a field of growing recent interest due to the promise of efficient signal acquisition, manipulation and recovery operations in applications where sensor utilization is a scarce, coveted, and/or expensive resource. Applications of CS technology on HIS acquisition and processing application promise to significantly increase the volume of relevant data captured in a single expedition and/or reduce the need of either continuously increasing available data storage resources or discarding valuable information. However, CS is an inherently lossy compression format, therefore it is necessary to evaluate the system parametrization that best fits the acquisition of data for a certain application. The current work shows that approaches based on this technology can improve the efficiency of several data processes, such as manipulation, analysis and storage, already established for hyperspectral imagery, without significant loss in data performance upon reconstruction. We present the results of a comparative analysis of classification performance between a HSI data cube acquired by traditional means, and HSI data cubes obtained through reconstruction from compressively sampled data points using a variety of different parameter combinations with a block-based CS recovery algorithm. To obtain a big picture view of the classification performance of compressively sensed HSI data cubes, we classify two verified HSI data cubes using a set of five classifiers commonly used in hyperspectral image classification. Global accuracy statistics are presented and discussed for each parameter combination, as well as class-specific statistical properties of the evaluated HSIs. Additionally, we perform a study of a state-of-the-art sparse-representation-based classifier(SRC) for hyperspectral imaging, and evaluate potential improvements that reduce computational complexity with comparable results. With the intention of furthering the extent of our classification performance analysis, we use these SRC-based classifiers in addition to the five conventional HSI classification algorithms considered. In this thesis, we show that using sampling factors below the values that allow for nearperfect HSI recovery has a positive impact on classification performance, and higher sampling factors do not have a negative impact on classification performance when compared to that of the original, uncompressed HSIs. Further, the use of large block sizes do not have a positive impact on classification performance that offsets the high computation time that comes with them. Finally, our proposed improvement to the SRC-based HSI classifier showed results comparable to a state-of-the-art approach, with the added benefit of reduced computational complexity.