Image Sequence Retrieval for Forecasting Weather Radar Echo Pattern

A novel framework and method are proposed to retrieve image sequences with the goal of forecasting complex and time-varying natural pattern. As a such pattern, we tackled weather radar images representing the spatial distribution of precipitation, the application is the local short-term forecasting of precipitation. In our approach, the forecast is made by retrieving past patterns similar to the present pattern, and the forecast pattern is produced by using its subsequent pattern of the retrieved sequences. In addition to the global distribution feature and velocity field of the pattern, we extract temporal texture features to catch the features of the echo patterns, which are nonrigid and deformable, and that appear and disappear. Similar sequences are retrieved based on a distance measure between paths in eigenspaces derived from the feature vectors. Several experiments confirm the performance of our retrieval scheme and indicate the predictability of the pattern.