To perform pyramidal analysis, reduced resolution versions
of such rain maps were built of size 80×80, 40×40 and
20×20. To speed up the whole procedure, which tends to
slow down considerably with larger maps, the maximum
size considered was 80×80.
Our improved technique proved to be generally slightly
better than the original, but worse, unexpectedly, for nextframe
forecasting. This was thought to be due to the socalled
local linearity phenomenon: parameter trend is far
from linear, but only when observed on intervals longer
than three frames. On intervals that short, linear forecasting
remains unbeatable. A summary with the most significant
results is reported in table 1. Values indicated here
represent mean square error over the entire image (of pointwise
differences: forecasted-observed), and are expressed in
terms of mm/h.