Memory-Based Forecasting for Weather Image Patterns

A novel method and a framework called Memory-Based Forecasting are proposed to forecast complex and timevarying natural patterns with the goal of supporting experts’ decision making. This paper targets the local precipitation phenomena captured as echo patterns in weather radar images, and aims to realize a tool that supports weather forecasters. In our framework, past image patterns similar to the present pattern are retrieved from a large set held in an image database, and the forecast image is produced by using the patterns that follow the retrieved patterns; it is analogous to human forecasters who imagine the future patterns based on their past experience. Appearance-based image features and temporal texture features are introduced to characterize the non-rigid complex echo patterns found in such radar images. The dissimilarity between two image sequences is defined as the normalized distance between paths of feature points in eigenspaces of the image features to retrieve similar past sequences. Forecast images are then constructed from a future point in the feature spaces, which is estimated by a nonlinear prediction scheme. Statistical experiments using weather radar images verify the effectiveness of our method and framework especially for drastically changing patterns.