A Genetic Algorithm (GA) tool for automated tuning of an operational cloud detection algorithm

The Air Force Weather Agency (AFWA) Cloud Depiction and Forecast System, version 2 (CDFS-II), ingests data from twelve meteorological satellites (GOES, METEOSAT, MT-SAT, TIROS, and DMSP), produces a cloud mask for every satellite transmission, and integrates these data into the hourly Worldwide Merged Cloud Analysis (WWMCA).

Every cloud mask product is the product of multiple cloud detection tests, each designed to detect a particular signature within the satellite imagery that helps to distinguish between clouds and cloud-free backgrounds. The cloud detection algorithm incorporates adjustable thresholds in the individual tests to adapt to seasonal variations and to eliminate false signals due to noisy transmissions or imprecise estimates of background conditions. These thresholds are unique to each satellite, and can vary between geographical regions and between different times of the day. There are thus hundreds of individual thresholds, many of which have interrelated effects. Maintaining this set of thresholds has proven to be difficult and time consuming. Extensive training is required to prepare someone for the task.

AFWA is developing a new tool employing a genetic algorithm engine to perform a large part of this tuning effort. For each tuning session, an experienced weather specialist must analyze one satellite transmission and produce a cloud mask that is as nearly perfect as possible. The genetic algorithm tool then operates upon a pre-selected subset of the thresholds to find the values that will produce an automated cloud mask matching the “perfect” mask as closely as possible.

The GA tuning algorithm is being implemented on CDFS-II in a way that parameter tuning may be performed by any professional experienced in satellite data interpretation or meteorology with only a few hours of setup work. The GA tool then runs largely unsupervised until it reaches an acceptable solution. This produces results that improve upon current processes while consuming fewer human resources. The GA tuning tool employs a standard GA class which is readily tailored toward solving the tuning problem. The GA tool is based on the existing “AutoGA” tool, which is built as a standard C++ class, and is dynamically scalable to the problem at hand. This paper will describe the genetic algorithm approach to the CDFS-II algorithm tuning process and its implementation into the existing cloud analysis process. We will show some initial results from the GA implementation, and how the GA technique improves the quality of the worldwide merged cloud analysis produced by CDFS-II.