Compressive Hyperspectral Sensing
A New Platform for Cheap, Efficient Hyperspectral Data Acquisition
A Smarter Hyperspectral Camera
It is a challenging task to design and develop a 2D hyperspectral instrument based on a single photon detector because we significantly under-sample spatially. We propose to use a spatial light modulator such as a DMD or LCD to selectively aggregate the light signals at different spatial locations to a single point as compressive samples. Furthermore, a grating is used to separate the incoming light into different wavelengths that can be detected by normal image sensors such as semiconductor charge-coupled devices (CCDs) or single photon detectors.
The Promise of Compressive Sensing
CCDs are widely available and inexpensive, but the acquisition time is fairly slow (e.g., 30 Hz). Although compressive sensing theory shows the number of compressed samples (m) can be an-order-of-magnitude smaller than that of the spatial samples (n), (i.e., m = O(klog(n)), where k is the sparsity or compressablitiy of the spatial signal in some transformed domains), for a 2D image that has 2.3 million pixels, we will need approximately 230,000 compressed samples, which will take a normal 30 Hz CCD more than 2 hours to produce one image. To this end, we propose to use a second spatial light modulator to selectively aggregate the light at different wavelengths into a single point. Finally, we propose to use single photon detectors, which can operate dramatically faster at MHz sampling rates and reduce 2D hyperspectral image acquisition time significantly. Furthermore, This design will reduce the required number of single photon detectors to one and thus reduce the cost of the instruments.
Essential Features of the Prototype Optical System
The proposed optical system must simultaneously filter incoming light from a scene both spatially and spectrally by a programmable spatial light modulator.
Compressive sensing gives us the opportunity to sample only enough of the data cube to deliver the information of interest. Delivering this may be done either by pseudorandom sampling of the scene, or much more efficiently by some adaptive algorithm that only samples the small portion of the data cube that is of interest. Since the portion of the data cube that will be of interest may consist of arbitrary combinations of spatial and spectral information, the system will provide arbitrary spatial and spectral filtering.
Inevitable tradeoffs between field of view, speed and , spectral resolution will be the most difficult decisions to make. The proposed device, in order to deliver substantial improvement in speed, will suffer in the other aspects. We may consider development of more than one optical system for different applications, where different tradeoffs will be struck.
Applications in Ecology and Agriculture
Our initial tests will be aimed at agricultural applications, including:
- Species identification and classification
- E,g, weed recognition
- Pattern recognition within species
- Different genotypes
- Crop groundcover
- Nitrogen condition of leaf
- Water stress effects
- Waxy leaves
- Also can use such cameras to analyse seed or grain as it comes into harvesters and predict its composition.