Monday, 23 January 2012: 2:00 PM
A Machine Learning Approach for Precipitation Estimation From Multiple Satellite Information
Room 242 (New Orleans Convention Center )
Accurate monitoring precipitation is important for flood forecasting and water resources system planning. Satellite remote sensing techniques provide a unique way for consistent precipitation observation at a global scale. In the development of satellite-based precipitation retrieval algorithms, both effective but compliment sensors and sampling capability from Low Earth Orbit (LEO) and geosynchronous (GEO) satellites should be considered. In this study, a method being capable of integrating multiple satellite sensors and platforms with effective data assimilation ability is developed. The proposed method includes several components: (1) a 2-D cloud tracking system, which captures the cloud advection from the subsequent GEO satellite images, (2) an empirical cloud motion and rainfall generation model, and (3) a sequential updating procedure using Kalman filter. In this algorithm, the GEO satellite images are used for tracking the cloud motion in time and the empirical cloud model is sued for rainfall estimation. An update scheme (Kalman filter) is used for near-real-time adjustment of state variables and rainfall estimates based on limited available rainfall estimation from LEO satellite passive microwave (PMW) sensors. With state variables sequentially adjusted from the PMW rainfall, the updated rainfall estimates are generated at the scale following the GEO satellite imagery. The model is evaluated over the CONUS using hourly radar rainfall data. The results show effective for rainfall estimation in sub-daily to hourly scale.