Towards the Concurrent Execution of Multiple Hyperspectral Imaging Applications by Means of Computationally Simple Operations

Abstract

The on-board processing of remotely sensed hyperspectral images is gaining momentum for applications that demand a quick response as an alternative to conventional approaches where the acquired images are off-line processed once they have been transmitted to the ground segment. However, the adoption of this on-board processing strategy brings further challenges for the remote-sensing research community due to the high data rate of the new-generation hyperspectral sensors and the limited amount of available on-board computational resources. This situation becomes even more stringent when different time-sensitive applications coexist, since different tasks must be sequentially processed onto the same computing device. In this work, we have dealt with this issue through the definition of a set of core operations that extracts spectral features useful for many hyperspectral analysis techniques, such as unmixing, compression and target/anomaly detection. Accordingly, it permits the concurrent execution of such techniques reusing operations and thereby requiring much less computational resources than if they were separately executed. In particular, in this manuscript we have verified the goodness of our proposal for the concurrent execution of both the lossy compression and anomaly detection processes in hyperspectral images. To evaluate the performance, several images taken by an unmanned aerial vehicle have been used. The obtained results clearly support the benefits of our proposal not only in terms of accuracy but also in terms of computational burden, achieving a reduction of roughly 50% fewer operations to be executed. Future research lines are focused on extending this methodology to other fields such as target detection, classification and dimensionality reduction.

Real-Time Hyperspectral Data Transmission for UAV-Based Acquisition Platforms

“…As previously described in this work, the HyperLCA compressor was developed following an unmixing-like strategy in which the most different pixels present in each image block are perfectly preserved through the compression-decompression process. This is traduced in the fact that most of the information that is lost in the compression process corresponds to the image noise, while the relevant information is preserved, as demonstrated in [20,29]. In [20], the impact of the compression-decompression process within the HyperLCA algorithm was tested when using the decompressed images for hyperspectral linear unmixing, classification, and anomaly detection, demonstrating that the use of this compressor does not negatively affect the obtained results.…”

“…This specific study was carried out while using well known hyperspectral datasets and algorithms, such as the Pavia University data set coupled with the Support Vector Machine (SVM) classifier, or the Rochester Institute of Technology (RIT) and the World Trade Center (WTC) images coupled with the Orthogonal Subspace Projection Reed-Xiaoli (OSPRX) anomaly detector. The work presented in [29] carries out a similar study, just for anomaly detection, but using the hyperspectral data that were collected by the acquisition platform used in this work and with the exact same configurations, both in the acquisition stage and compression stage. Concretely, the data used in this work, as described in Section 4.2, are a reduced subset of the hyperspectral data used in [29].…”

“…The work presented in [29] carries out a similar study, just for anomaly detection, but using the hyperspectral data that were collected by the acquisition platform used in this work and with the exact same configurations, both in the acquisition stage and compression stage. Concretely, the data used in this work, as described in Section 4.2, are a reduced subset of the hyperspectral data used in [29].…”

“…In particular, a smaller portion of 825 subsequent hyperspectral frames of 1024 pixels and 160 spectral bands each was selected for the experiments (as highlighted in Figure 7c. Despite the Specim FX10 camera used for acquiring this data can collect up to 224 bands, the first 20 bands and the last 44 ones were discarded during the capturing process due to a low Signal-to-Noise-Ratio (SNR), as described in [29]. Using the described hyperspectral image, an acquisition simulation process has been developed, which independently stores each of the 825 hyperspectral frames as a single file in a target directory and at user defined speed.…”

FPGA-Based On-Board Hyperspectral Imaging Compression: Benchmarking Performance and Energy Efficiency against GPU Implementations

“…Additionally, the HyperLCA compressor provides quite satisfactory rate-distortion results for higher compression ratios than those achievable by lossless compression approaches. Furthermore, the HyperLCA algorithm preserves the most characteristic spectral features of image pixels that are potentially more useful for ulterior hyperspectral analysis techniques, such as target detection, spectral unmixing, change detection, anomaly detection, among others [30][31][32][33][34]. Figure 1 shows a graphic representation of the main computing stages involved by the HyperLCA compressor.…”