Genetic refinement of cloud-masking algorithms for the Multi-spectral Thermal Imager (MTI)

The Multi-spectral Thermal Imager (MTI) is a high-performance remote-sensing satellite designed, owned and operated by the U.S. Department of Energy, with a dual mission in environmental studies and in nonproliferation. It has enhanced spatial and radiometric resolutions and state-of-the-art calibration capabilities. This instrumental development puts a new burden on retrieval algorithm developers to pass this accuracy on to the inferred geophysical parameters. In particular, the atmospheric correction scheme assumes the intervening atmosphere will be modeled as a plane-parallel horizontally-homogeneous medium. A single dense-enough cloud in view of the ground target can easily offset reality from the calculations, hence the need for a reliable cloud-masking algorithm. Pixel-scale cloud detection relies on the simple facts that clouds are generally whiter, brighter, and colder than the ground below; spatially, dense clouds are generally large on some scale. This is a good basis for searching multispectral datacubes for cloud signatures. However, the resulting cloud mask can be very sensitive to the choice of thresholds in whiteness, brightness, temperature, and connectivity. We have used a genetic algorithm trained on (MODIS airborne simulator-based) simulated MTI data to design a cloud-mask. Its performance is compared quantitatively to hand-drawn training data and to the EOS/Terra MODIS cloud mask.