A new method, based on an analysis of radar reflectivity images, is proposed for the short-term forecasting of precipitation distributions. Based on the repeatability of weather phenomena, past patterns similar to the present pattern are retrieved from a database storing a large set of radar images, and used to produce forecast information based on the temporal development of the retrieved patterns.
So far, a number of extrapolation methods have been applied to forecast short-term precipitation, but the methods fail to produce reasonable forecasts especially for drastically changing patterns, because of their assumption on the persistency of phenomena. The approach of using similar past patterns, which allow deterministic non-linear predictions of chaotic time series, has the potential to predict such pattern changes as from one pattern to another.
However, it has been difficult to represent two dimensional complicated motion pattern concisely, and only a few studies have tackled the forecasting of two dimensional patterns based on this approach. To solve the problem, we have developed an original feature extraction scheme. The proposed features include so called temporal texture feature, a mesh feature, and a velocity field. The temporal texture feature can characterize types of precipitation structures such as stratiform, convective pattern, band-shaped pattern, and scattered pattern. The mesh feature represents the spatial distribution of echo patterns as regards global position and pattern shape in an image plane. The velocity field indicates the spatial distribution of atmospheric flow at the radar scan altitude.
Using the extracted feature vectors, the echo pattern at a particular time is transformed into a point in the feature space whose axes are elements of the feature vector, and a sequence of patterns is represented as a path in the feature space. Here, we compute a dissimilarity measure between two patterns as the distance between path segments in the feature space. Based on the dissimilarity measure, we retrieve matched past patterns with small dissimilarity from the database which includes pre-calculated feature vectors corresponding to the image sequences. The patterns that follow the retrieved patterns are used to create forecast information.
In our system, retrieved past sequences and subsequent sequences are displayed to users e.g. meteorologists, in order to support their decision making. Also these sequences are used to synthesize the forecast images representing possible precipitation distributions in the near future. Several experiments are conducted to confirm the feasibility of the proposed method using PPI radar images collected in Sapporo area, Japan, over three winter seasons