Tuesday, 8 January 2013: 11:30 AM
Room 11AB (Austin Convention Center)
Meteorologists and other scientists rely heavily on remotely sensed data collected from instruments aboard orbiting satellites. The design of such instruments requires technical and economic trade-offs that results in certain desirable data not being directly available. One way to mitigate the lack of availability of this data, is to use machine learning techniques to estimate the data that is not directly observed. This can be accomplished by exploiting statistical correlation with information in available data sets. By combining the information from multiple other sources it is often possible to create an accurate estimate of the physical parameters which are not directly observed. We apply this idea toward the problem of estimating the 13.3 micron band for the Visible Infrared Imaging Radiometer Suite (VIIRS), an instrument aboard NOAA's operational satellite, Suomi NPP. The radiance from the 13.3 micron band is not directly available from VIIRS although this band has important applications such as estimating cloud-top pressure. We demonstrate that a reliable estimate of this band can be made using other VIIRS bands such as M16 (12.01µm), M15 (10.76µm), M14 (8.55µm), and M13 (4.05µm), as well as input from the Cross-track Infrared Sounder (CrIS), which produces data at much finer spectral resolution, making measurements in hundreds of nearby infrared bands, though with lower spatial resolution. In order to establish feasibility and develop our estimation methods, we used data from the Moderate Resolution Image Spectroradiometer (MODIS), an instrument aboard the NASA satellites Aqua and Terra, part of the Earth Observing System (EOS). This instrument has bands that match those of VIIRS. In particular, MODIS bands 32, 31, 29, and 23 have characteristics similar to bands of VIIRS. In addition the MODIS instrument band 33 is centered at 13.3 microns, which is the target spectral band, and has the same resolution, location, and temporal characteristics as desired. To stand in for the CrIS component, we used data from the Atmospheric Infrared Sounder (AIRS), which is also onboard the Aqua satellite. Like CrIS, AIRS covers the target spectral response range around 13.3 microns (through a multitude of narrow bands), and like CrIS at a much lower spatial resolution. The working assumption we test in this work is that there is a function which can map radiances from bands that will be available on VIIRS to estimate the target 13.3 micron band. To fit this function we will assume a measure of scale-invariance in the relationship between available source bands and the target band. Thus that if we establish the relationship between a vector of radiance values in the source bands, and the scaler radiance value in the target band, using low resolution data, being a point-wise relationship, we assume it to hold at higher resolution. For computing the relationship we build a database of source and target values using our CrIS proxy AIRS (low resolution). These are created for the target 13.3 micron band by using the known spectral response function of MODIS band 33 and applying weighted averaging to 109 corresponding narrow-response AIRS bands, which has been shown to reproduce MODIS band 33 at the lower AIRS spatial resolution to an extremely high degree of accuracy. To make corresponding pixels for the other MODIS bands, the high-resolution images were degraded to the AIRS spatial resolution by averaging those values which are actually in each AIRS pixel's field of view. The estimation function of the 13.3 micron band is implemented using a k-nearest neighbour search, and locally averaging the results. In particular, to estimate the target radiance at a given pixel, the corresponding vector of radiance values for the source bands at high resolution is used to query the database. The query is efficiently executed using the k-d tree data search algorithm to find k-nearest neighbours. The corresponding target band 33 values for these neighbours are then averaged (using Gaussian weighting based on distance in the data space) to create an estimated value for each pixel at the higher resolution in the target band 13.3 micron band. In addition to testing the 13.3 micron values produced by our estimation against the known MODIS band 33, we tested it as input values to an algorithm which estimates cloud top pressure using data from 11, 12, and 13.3 micron bands. This algorithm was developed for the Geostationary Operational Environmental Satellite R Series (GOES-R). That satellite will be launched as soon as 2015 and will carry the Advanced Baseline Imager (ABI) which will measure 13.3 micron band (though at a lower 2km spatial resolution). These tests showed that similarly-synthesized data from VIIRS and CrIS would allow VIIRS/CrIS to match GOES-R in terms of cloud-top pressure determination, to within the GOES-R specifications, which is especially important for getting such values for night scenes since GOES-R, unlike VIIRS, relies on data in the visible to near-infrared range.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner