Multi Sensor Data Fusion

Storm cells resulting from convection are one of the primary objects of interest in meteorology. These are indicated by thick clouds that extend high in the atmosphere. Since pressure decreases with altitude, the height of the cloud can be estimated by the cloud top pressure. The cloud-top pressure product critically uses 13.3-μm band, which is available as band 26 on the MODIS instrument. The Visible Infrared Imaging Radiometer Suite (VIIRS) on NOAA’s Suomi NPP satellite is tasked with monitoring the earth, but unfortunately it does not have a 13.3--μm band.

The goal of this work is to estimate a 13.3-μm band for VIIRS by fusing VIIRS data with another instrument on Suomi NPP: an infrared sounder. The Cross-track Infrared Sounder (CrIS) on Suomi NPP has 1305 narrow band spectral channels in the infra-red. The differing properties of the atmospheric constituents across these bands makes it possible to estimate, for instance, temperature and humidity at different altitudes in the atmosphere. Fortunately for the algorithm development and evaluation there is an instrument similar to CrIS, the Atmospheric Infrared Sounder (AIRS), on Aqua, which carries MODIS. The sounders CrIS and AIRS both have a much lower spatial resolution (and somewhat smaller field of view) than the imagers VIIRS and MODIS. The span of the spectral bands on each sounder can easily be used to create a 13.3-μm band, but at much lower spatial resolution than the native VIIRS bands. The goal is to estimate a VIIRS native resolution 13.3-μm band.

The first step is to create the lower resolution 13.3-μm band estimated from the sounder data. Those imager bands which are common to both MODIS and CRIS are selected as inputs for a 13.3-μm regression. The imager bands are first down-sampled to match the resolution of the low resolution 13.3--μm band. The basic idea is to learn a function which takes a tuple of values at a pixel in the input bands and outputs the 13.3-μm value. The with the downsampling of VIIRS and the information from CrIS the training data for this function is available low resolution. If we simply apply the function we learned pixel-wise to the high resolution input bands in VIIRS, we can produce a high resolution estimate of 13.3-μm . We evaluated this using the MODIS-AIRS data and found that this estimate is not very accurate because while locally the relationship between the input bands and the 13.3-μm band is a function, this relationships varies across the image. To account for this spatially varing relationship, we implement the mapping as a search that takes each high-resolution tuple of values in the input bands, along with the relative x-y location in image coordinates as inputs for the search and use a nearest neighbor lookup which has a term that favors matches which are close in the image. This local non-parametric regression estimates a 13.3-μm band was evaluated by the MODIS-AIRS pair of instruments, where the high resolution 13.3-μm can be used for validation. We then apply this to the VIIRS-CrIS pair to estimate the 13.3-μm band where it is then used for cloud-top pressure estimation.