Application of Shapley-fuzzy neural network method in long-time rolling forecasting of typhoon satellite image in South China

In order to make better use of the observational data of a large number of satellite cloud images to improve the forecasting ability of typhoon rain, increase the forecast accuracy of cloud system changes related to typhoon heavy precipitation, and increase the timeliness of the forecast for future cloud system changes, a model is constructed based on the cooperative strategy Shapley-fuzzy neural network for non-linear intelligent calculation of satellite cloud images of typhoons in South China. It is of great significance to enhance the practicability and timeliness of satellite cloud image data in the weather forecast of typhoon rain. Based on satellite cloud images of typhoon processes in South China during 2013—2016, an ensemble forecast method similar to the numerical forecast model is used to decompose the sample sequence of the cloud top brightness temperature of the satellite cloud images at 6 h intervals by the empirical orthogonal function, and the time coefficients decomposed are used as the forecast component in the cloud image forecast model. The development and changes of typhoon clouds are mainly affected by the atmospheric environmental flow field. Here, the physical quantity forecast products of the numerical forecast model are used as the forecast factors of individual forecast components of cloud images. And the stepwise variable selection algorithm of k-nearest neighbor mutual information estimation is adopted to realize the selection of related variables and the elimination of weakly related variables through a two-step process. The Shapley-fuzzy neural network ensemble forecast model of the corresponding time coefficient components is established, and the time coefficients and space vectors of each cloud image obtained by the forecast are further synthesized to reconstruct the forecast cloud image at the future time. In this way, the long-term and objective rolling prediction of the satellite cloud image at forecast lead time from 6 h to 72 h is realized. The prediction results of satellite cloud images of typhoons in southern China show that the cloud images predicted by the new scheme are highly correlated with the observed cloud images. The reconstructed predicted cloud image's basic contour, texture feature distribution, clarity, and cloud strength are relatively close to that of the observed cloud image. In addition, based on the forecast factors of the same cloud images, this paper further used the multiple linear regression scheme to predict the cloud images consistent with the new scheme for the same modeling and forecasting samples. Comparison of the results shows that this nonlinear forecasting model can better predict long-term development, movement, and trend of typhoon cloud cluster than the linear scheme. And the predicted cloud images are more consistent with the main features of the observed cloud images. The valid time of the cloud image forecast in this experiment has reached 72 h, which has a practical significance in operational weather forecasting.