Synthetic satellite visible imagery

Although numerical weather forecast models can provide high-resolution forecasts up to a day ahead, it can be hard to visualize what the forecasts actually mean because the forecast variables -- pressure, mixing ratios, etc. are not what someone thinks about when they think "weather". One way to address this problem is to create synthetic radar and satellite images from the model forecasts. Synthetic radar images are common (for example, see here) and very useful for forecasts of convection. But what about simple cloud cover? For that, you'd need to simulate satellite visible images. And those turn out to be very hard. It takes tremendous processing power to create a synthetic satellite visible image.

One of the things I worked on in the past year or so is to create a statistical approximation to the radiative transfer model that is used to create synthetic visible images. You can see the synthetic visible image for today's forecast run (essentially a prediction of the weather until tomorrow) here (click on "US").

p.s. One thing that's readily obvious from that loop is that our statistical model incorporates the sun angle, so the synthetic satellite image darkens over time (real satellite visible images are not available at night). Obviously, this is pointless -- we'll have to create a training set that doesn't incorporate terminator lines. This is harder than it sounds.

p.s.2: Our statistical model approximates the transfer model, not the real data because model forecasts have significant position errors. This means that you can't train a pixel-to-pixel transfer function.