The Method Of Typhoon Precipitation Nowcasting In Considerision Of Large-scale Complex Spatio-temporal Patterns

Typhoon precipitation has been a serious threat to economic development,production and life at home and abroad.Real-time and accurate typhoon precipitation nowcasting plays an important role in disaster prevention and mitigation.Due to the complex spatiotemporal evolution pattern and the specific cause and performance,typhoon precipitation is different from ordinary precipitation and has more complex spatio-temporal nonlinear characteristics,which puts forward higher requirements for spatio-temporal model mining and nowcasting.The traditional numerical weather prediction(NWP)model has some shortcomings in short-term prediction due to its large time cost and high computation ability requirement.Deep learning has obvious advantages in short-term prediction due to its powerful nonlinear fitting ability,flexibility and rapidity.However,deep learning currently focuses on the general precipitation nowcasting,and has not been deeply studied and effectively applied in the field of typhoon precipitation nowcasting.The problems that how to obtain large-scale high-precision near-real-time precipitation data,integrate typhoon and precipitation multi-modal data effectively,model dynamic spatio-temporal correlation of local typhoon precipitation,capture long-distance spatiotemporal dependence that characterizes higher-order spatio-temporal characteristics of macro weather system,mine complex dynamic spatio-temporal model of large-scale typhoon precipitation,and construct an accurate and real-time nowcasting model of typhoon precipitation,are the direction of deep learning research to be further studied and discussed.Based on high spatio-temporal resolution near-real-time meteorological satellite images,we preliminarily established a deep learning method for typhoon precipitation nowcasting considering large-scale complex spatio-temporal patterns.Taking typhoon precipitation events in the northwest Pacific as an example,we carried out method research,experimental verification and application analysis.The main contents are summarized as follows:(1)Based on infrared remote sensing and precipitation data,the pre-processing of the multi-source data,such as spatio-temporal matching,outlier processing,data screening and data set division,was carried out.After that,the single-channel and multi-channel combination correlation analysis was carried out based on the field of view points of different precipitation intensities.At the same time,combined with channel attention mechanism,infrared channel contribution analysis was performed on different precipitation intensity recognition tasks in consideration of spatial pattern,and determined the key channels of infrared remote sensing precipitation estimation.(2)Aiming at the problem of precipitation data imbalance in large-scale scenarios,a multi-task collaborative data balance framework for remote sensing precipitation estimation was proposed to solve the error transfer problem of existing two-stage framework solutions.By designing multi-task consistency constraints and feature-level cross-branch interaction modules,a positive information feedback loop was formed among tasks.The model’s preference for medium and high intensity precipitation was gently improved,while the negative impact of error transfer between tasks was suppressed,and the accuracy of large-scale remote sensing precipitation estimation was significantly improved.(3)Based on how to construct typhoon representation,integrate multi-source heterogeneous data of typhoon precipitation,and mine dynamic spatio-temporal evolution pattern of local typhoon precipitation,combined with the typhoon track intensity information,we proposed a typhoon feature map construction scheme and verify the effectiveness of the multi-branch fusion framework by comparison.At the same time,a dynamic graph-guided spatio-temporal modeling mechanism for local typhoon precipitation was proposed to enhance the regional feature expression of typhoon precipitation and improve the model’s nowcasting ability of high intensity precipitation.(4)In order to consider the macroscopic spatio-temporal evolution model of largescale weather system and carry out long-distance spatio-temporal modeling,we proposed a long-distance spatio-temporal modeling mechanism of precipitation based on cross-patch multi-layer semantic attention.By aggregating neighborhood information and condensing redundant fine-grained information,long-distance spatio-temporal modeling was carried out based on multi-scale cross-patch spatio-temporal features with full spatio-temporal vision,which improved the model’s perception ability of higher-order spatio-temporal features of macroscopic weather systems at less computational power cost.On the basis of understanding and mining the temporal and spatial patterns of typhoon precipitation,we realized the deep learning method innovation of typhoon precipitation nowcasting,significantly improved the accuracy of typhoon precipitation nowcasting model,supported the application of typhoon precipitation nowcasting based on remote sensing in large-scale scenarios,and promoted the cross research and integrated development of remote sensing pattern recognition,artificial intelligence and extreme weather.