Real-Time Atmospheric compensation and surface temperature estimates from Satellite using Neural Networks

Military satellites are being used to determine potential security threats using accurate high sensitivity measurements from defense satellites. These systems focus much more on sensitivity and Signal/Noise Ratio (SNR) than on multispectral capability. Therefore, these systems are highly vulnerable to thermal absorption and reemission that can affect the thermal signature prior to launch. In particular, there is a need to use additional assets to estimate the atmospheric temperature and water vapor profile so that an estimate of the atmospheric processes can be obtained and a correction developed to improve the thermal detection. In particular, the use of existing meteorological geostationary assets is crucial to this effort since these satellites can estimate the atmospheric temperature and water vapor profiles which in principle can be “inverted” to get the surface signature. However, this approach cannot be implemented in real time efficiently and therefore we need to develop a more empirical compensation approach which can be used in real time with minimum computer resources. In this thesis, we develop a Neural Estimator approach to take metrological inputs of water vapor and temperature over three integrated pressure levels WV1 ( 0.9Ps \u3c P \u3c Ps), WV2 (.7Ps \u3c P \u3c 0.9Ps) , and WV3 (.3Ps \u3c P \u3c .7Ps) from the existing GOES-13 sounder with additional information like zenith angle, and radiances processed from MODTRAN. The result was a robust neural network that could be applied to multiple sites and weather scenarios without much error in the ground temperature outputs. With an RMSE of .90 (K) in comparison to surface temperatures, good correlations on a near real time feed could be possible by this approach to give detection targets and provide instantaneous results to ground temperature unknowns