Presented by: Chung-Hsien Yu Spatio-Temporal Asynchronous Co-Occurrence Pattern for Big Climate Data towards Long-Lead Flood Prediction Department of Computer Science Presented by: Chung-Hsien Yu Department of Civil and Environmental Engineering 2015 IEEE International Conference on Big Data
5-Day Precipitation Forecast Error Amplification Heavy Rainfall Flooding Source: http://www.wpc.ncep.noaa.gov/ensembletraining/ 2015 IEEE International Conference on Big Data Paper ID: BigD488
Problems and Solutions Simulation Data Mining Short lead-time: Error Amplification Long lead-time: Source Location Backtracking Global scale model: Entire spatial features Computationally expensive Local scale model: Feature reduction Target location focus 2015 IEEE International Conference on Big Data Paper ID: BigD488
Source Location Backtracking How to find the associations? Temporal Clusters Heavy Rainfall How to extend the lead-time? Flooding time series 2015 IEEE International Conference on Big Data Paper ID: BigD488
Spatio-Temporal Asynchronous Co-Occurrence Pattern Source location A Asynchronous Co-Occurrence Locations Lead-time l Target location Source location B C Source location 2015 IEEE International Conference on Big Data Paper ID: BigD488
Asynchronous Co-Occurrence Locations How to find the asynchronous co-occurrence locations? length(PWC) 8+3 PWC A Source location Lead-time=7 EPC 3 Co-Occurrence Band=3 Target location Co-Occurrence Rate=3/11 The locations with long PWC… Asynchronous Co-occurrence Locations: The locations with over average COR and high COB. 2015 IEEE International Conference on Big Data Paper ID: BigD488
Spatial-Temporal Predictive Modeling Ensemble model of different lead-time Asynchronous Co-occurrence Locations Monitoring the formation progresses l=4 latitude longitude l=5 l=6 2015 IEEE International Conference on Big Data Paper ID: BigD488
The Proposed Framework Big Climate Data 2015 IEEE International Conference on Big Data Paper ID: BigD488
Case Study: Flood Prediction in Iowa Avg. time: 15 sec Number of features: 9 5,328 9 500 Avg. time: 187 sec Predict the severe floods caused by heavy precipitation at least 9 days in advance! Atmospheric factors Gridded locations 2015 IEEE International Conference on Big Data Paper ID: BigD488
Summary Associating precipitation and precipitable water for long-lead flood prediction. Reducing the training data size for modeling. Map-Reduce-style parallelization. Scalable 2015 IEEE International Conference on Big Data Paper ID: BigD488
Acknowledgments Chung-Hsien Yu csyu@cs.umb.edu Prof. Wei Ding Dong Luo Joseph Cohen Prof. Shafiqul Islam Prof. David Small Chung-Hsien Yu csyu@cs.umb.edu Research Computing Department at the University of Massachusetts Boston 2015 IEEE International Conference on Big Data Paper ID: BigD488