The Radiometric Stability of VIIRS/SNPP RSB over Lake Michigan Using a Machine Learning Technique

The Visible Infrared Imaging Radiometer Suite (VIIRS) is a key instrument carried on the JPSS satellite series. It is critically important to understand the VIIRS long-term radiometric stability through the time series analysis. Our previous studies have been focused on regions with deep convective clouds (DCC) and lakes with stable surface such as Salton Sea in California. For other regions like Lake Michigan, however, it is difficult to monitor the radiometric stability of VIIRS directly through time series analysis from observations because of relatively strong temporal variations of surface properties due to biological activities and ice and the atmosphere above surface. Therefore, we proposed an alternative method to estimate the radiometric stability of VIIRS visible bands by taking into account the spectral correlations among different bands. This is based on the fact that even if there exist temporal variations on the surface and atmosphere above it, the spectral correlations remain stable in that all bands are observing the same targets. In this study, a machine learning framework is developed to learn the spectral correlations, i.e. predicting one specific VIIRS band from observations of the other 10 visible bands. The models are trained using observations from 2013 and found to have great prediction accuracy. Then the trend of prediction bias for each band is mathematically calculated. While the trend of prediction bias can be regarded as the trend of model prediction minus trend of observation, the model prediction trend is affected by observation trends from the other 10 bands. Sensitivity experiments are carried out with permutations to each predictor band of the predicting models to acquire approximate linear sensitivity coefficients for each predictor band to each target band. The model prediction trends can be approximated by the sum of these coefficients multiplied by their respective observation trends. Therefore, one can get a matrix of 11 equations with 11 unknown variables of observation trends. With the trends of prediction bias calculated from predictions and real observations, one can evaluate the stability of VIIRS bands from the solution of the observations trends. Results over Lake Michigan will be presented. This methodology once proved loose the requirement of stable homogeneous surface, and thus can be applicable to other surface conditions not limited to Lake Michigan.