Atmospheric Cloud Forecasting in Support of Space Based Applications

摘要

Many space based applications from imaging to communications are impacted by the atmosphere. Atmospheric impacts such as optical turbulence and clouds are the main drivers for these types of systems. For example, in space based optical communications, clouds will produce channel fades on the order of many hundreds of.decibels (dB) thereby breaking the communication link. Optical turbulence can also produce fades but can be compensated for by adaptive optics. The ability to forecast the current and future location and optical thickness of clouds for spaced based to ground optical communications is therefore critical in order to achieve a highly reliable system. We have developed an innovative method for producing such forecasts. These forecasts are intended to provide lead times on the order of several hours so that communication links can be transferred from a current clear ground location to another more desirable ground site. This modeling system is referred to as the Cloud Propagator Forecast (CPF) and it operates on successive, satellite remotely sensed, cloud analyses. The CPF produces probability forecasts of future cloud cover conditions at each point location. The forecasting algorithm is a combination of empirical Lagrangian and Eulerian regression over multiple spatial scales, but treats -time auto-regressively. Input cloud masks are transformed into proxies first. A cloud cover proxy is a variable which has a more Gaussian distribution than literal cloud cover. For a given pixel, the cloud cover proxy is computed first by determining whether at the initialization time the pixel was clear or cloudy. Clear pixels will be assigned only positive proxies; cloudy pixels will be given only negative proxies. The degree the assigned proxy is different than zero depends on the fraction of pixels in a small neighboring space which have similar cloudy/clearness. The neighboring space is approximately the spatial scale of a skydome and has a temporal scale of one hour. Pixels which are unlike their neighbors will have proxies close to zero, those largely identical to their neighbors will has proxies close to plus or minus one. Final cloud proxies are computed using a non-linear transform to stretch out the extremes into a pseudo-Gaussian distribution. The model then decomposes the proxy fields into scale-filtered components. Longer spatial scale patterns are expected to be more predictable over time;.shorted scales less so. Differentiating them allows the model to retain the maximum predictive skill through training. The resulting forecasts have several desirable characteristics. First, they evidence substantial skill when compared to persistence. Additionally, these forecasts extrapolate movement of cloud features, and also allow for degradation of fine scale features without compromising more predictable larger scales. The forecasts are reliable, in that.specific probability categories will assess at their stated probabilities, and also consequently unbiased. Details of the algorithm and results used for a realtime spaced based application will be shown at the conference.