1 of 6

HyperBand

Hyperspectral Edge Computing

Team 94 – NASA Space Apps, Sydney

2 of 6

What does HyperBand do?

  • In short, it converts hyperspectral data into multispectral data
  • Hyperspectral sensors have vast utility, but are costly to run due to huge data file sizes
  • Multispectral sensors are popular as the files are smaller, but they are hard-wired for specific applications
  • By using a hyperspectral sensor to produce multispectral data, you have essentially developed a reprogrammable multispectral imager, providing the best of both systems

3 of 6

The Data

  • The data used is from 1/1/2016, taken by the Earth Observing-1 NASA satellite with it’s Hyperion imaging spectrometer, publicly available on Earthdata
  • The image is formatted to align north, so a large amount of black space is in the image. This was ignored given the timeframe
  • The image is also presented in x-y tif tiles at different wavelengths. This was converted into x-λ images to better simulate the output from a pushbroom sensor before processing

The image in RGB (left) and an example single band at 1790nm (right, enhanced for viewing)

4 of 6

The Input

  • The input data was 4581 tif files giving slices in x-λ (242 by 1311 pixels), scanning down on the images seen before
  • When placed in sequence, the illuminated spectrum above moves from left to right. This is a product of the black space in the image

Slice 1816 of 4581, showing the spectral signature of the illuminated area

5 of 6

Inputdata.txt

  • The size of the image (y,x)
  • The sentinel bands. These are thin bands so only one input band is required to each output
  • A red and NIR band for NDVI. These are wider bands and so are presented as ranges. This could be condensed into one band by another algorithm before transmitting

6 of 6

Processing

Observations

  • In total, all 6 million pixels where processed into 10 bands in ~1hr50min
  • When tested, this time was found to be roughly linear to number of bands outputted
  • A lot of that time was spent accessing data which was done over USB or SD card on the Raspberry Pi

Recommendations

  • Speed needs to be increased if real-time reduction is the goal
  • Testing the hardware on a Jetson Nano or a SpaceEdge would give a better idea of if we’re at that stage
  • Faster storage hardware would increase the speed of the algorithm
  • File formats other then geotif may be needed both for compatibility and for speed