We have a 300 x 300 element data set that represents some geographic area, where each element/pixel is about 1km on a side. Each element holds one of two values: clear or cloud. The data is represented as a 2D array of 8-bit values. The data set is updated every 10 minutes, and we already have about 5 months of historical data.
This data can be considered "observed data" or truth data, and we wish to use this data to forecast future situations. Specifically, what sorts of ML algorithms/techniques could be used to forecast the future position of the cloud pixels in the dataset at times t=1, 2, ...n? I've been working on an algorithm that utilizes something like an SVD, finding a linear transformation from the t=-1 and the current (t=0) data set, and then applying the transformation to the current data set to predict where the elements will be at t=1, 2, n. A very early/rough implementation shows that it is very computationally expensive (which isn't a deal breaker, but is of moderate concern), and it seems to not perform that well when comparing the predicted positions of cloud pixels back to the truth/observed data that finally comes into the data set.
Any ideas for alternative approaches? Anyone know of any similar problems? White papers, books? Cheers!
[link][comment]