|
|
|
Adaptive Algorithms For Optimal Classification and Compression of Hyperspectral Images
|
|
|
| Hyperspectral images are obtained from large focal plane spectrometers with an array of multiple sensors that are arranged in an NxN matrix. Each sensor yields an Mx1 spectrum, where M is the number of image bands, typically between several dozens to several hundreds. Spectral and hyperspectral images are used today in most space science and earth science missions because of the richness of information they capture about the compositional properties of surface materials. Since hyperspectral images constitute very large data sets, they need to be compressed for storage and transmission. Classification of spectral species is the primary task in the analysis of hyperspectral imagery, with the goal of distinguishing as many species as possible, as accurately as possible. Therefore, a measure of good compression should be based on how well it preserves class distinctions. This, in general, does not correspond to a minimum distortion measure.
The goal of this project is "classification-driven compression" of hyperspectral images. That is, we propose to design and implement compression algorithms that will be optimized for efficient classification. The compression algorithms will be recursively optimized as a function of certain classification (or unsupervised clustering) metrics. Performance metrics will be defined for the classification schemes and some others for the compression algorithms. A combination of these metrics will be used to design a cost function, which in turn will be optimized to update the compression algorithms. In other words, the performance of the classification algorithms will dictate the real-time adaptation of the digital filters used in the compression schemes. This is a new concept, where the filters in the compression algorithms, instead of acting independently, form a closed loop coupled system with the classification (or clustering) algorithms. The algorithms will be designed for real-time on-board processing. Therefore, we seek to develop fast filtering algorithms with the lowest possible computational complexity. This work will change the way we look at and understand hyperspectral data. Since we will measure the quality of the compression by the retention capability of the distinction among spectral species, for any required compression ratio, we will much more successfully preserve the meaning of the data inherent in and specific to spectral images, than distortion based compressions. The proposed work was inspired by and will build on the results of two previous AISRP grant projects.
All space science missions that carry spectral imagers can significantly benefit from the results of this project in scenarios where lossy data compression is needed at compression ratios as high as can be achieved without loss of relevant spectral information. Because of our approach to compression, we will be able to determine the optimal compression ratio in an adaptive manner. The fast parallel implementation of the compression and classification algorithms, makes the combined system described in this proposal an intelligent, real-time on-board data understanding, compression and classification machine that learns on-chip and adapts continuously to new circumstances as desired, modifying the compression scheme to best suit a given environment. |
|
|
