1. An artificial neural networks system is used as model to estimate precipitation at 0.25° x 0.25° resolution. Two different networks are being developed, one for rainfall, and one for snowfall. The following steps are used to prepare the ANN model: 1) Snowfall a. Calculate ΔSWE using the difference in SWE between two consecutive days. b. Distinguish between rainy and snowy pixel by selecting only pixels with positive ΔSWE and temperature less than 0  C. c. Choose half of the pixels obtained from the previous step for calibration (train) of the ANN and the other half for validation of the model. 2) Rainfall a. Select only pixels with positive rainfall measurement in NEXRAD Stage IV data. b. Fifteen rainy hours were chosen for calibration of the ANN and another fifteen for validation of the model. Each ANN model was tried for GOES IR alone, and for combinations of IR with AMSU MW to test the improvement of precipitation estimates. From those tests, AMSU- MW 89 GHz and 150 GHz showed more correlation with precipitation, therefore these frequencies were selected to be used for precipitation estimation. National Oceanic and Atmospheric Administration - Cooperative Remote Sensing Science and Technology Center Abstract Estimating accurate precipitation from remote sensing information is still a challenge, particularly, for remote and mountainous regions where ground-based gauge networks and radar coverage are not available. Therefore, the main objective of this study is to develop a multi-sensor algorithm based on an artificial neural networks system for precipitation estimates using IR from Geostationary Operational Environmental Satellite (GOES) in conjunction with microwave from Advanced Microwave Sounding Unit (AMSU). Remotely sensed infrared provides brightness temperature only from the cloud top, but microwave spectrum can penetrate deeper and provide some information properties from inside the clouds. Consequently, using multi-sensor IR and MW information is expected to improve the accuracy of precipitation estimates. In this study, two types of precipitation are considered, rainfall and snowfall, and as a result, two different ANN models are in development. For rainfall, NEXRAD stage IV data is used to train and validate the rainfall ANN model, while the difference in two consecutive days of snow water equivalent (SWE) data from SNOTEL stations is used to train and validate the snowfall ANN model. Different storms were selected over the western United States, warm season storms are considered for rainfall and winter storms for snowfall. Preliminary investigation indicates that higher frequency microwave (AMSU-89 GHz and -150 GHz), is more correlated with precipitation, then it is an appropriate source of information to be combined with the GOES infrared channel 4, and used as model input for precipitation estimation. For validation data, the rainfall model estimates using combination of MW and IR is more correlated with NEXRAD stage IV, with average correlation coefficient of 0.25 (about 0.19 for AMSU-89 GHz and, about 0.31 for -150 GHz), than using only IR as model input, with an average correlation of about 0.14. The same occurs with the snowfall model estimates, the combination of MW and IR is more correlated to ΔSWE with average correlation coefficient of 0.44 (about 0.36 for AMSU-89 GHz and, about 0.52 for -150 GHz), than using only IR as model input, with an average correlation of about 0.32. However, the ANN networks needs to be modified to improve those precipitation estimates. Using Multi-Spectral Satellite Based Information for Precipitation Estimation Cecilia Hernández, Dr. Shayesteh Mahani, and Dr. Reza Khanbilvardi NOAA Collaborator: Ralph Ferraro (CICS) & Dr. Robert Kuligowski (NESDIS) Email: chernandez-aldarondo@gc.cuny.edu, mahani@ce.ccny.cuny.edu MethodologyPreliminary Results The preliminary results presented here are from the best ANN obtained so far, using backpropagation and adaptative learning. The relationships between model precipitation estimates vs. independent observations showed an average correlation coefficient of about 0.14 and about 0.32, using only IR for rainfall and snowfall, respectively. Moreover, the average correlation coefficient obtained using the combination of MW and IR was about 0.25 for rainfall, and about 0.44 for snowfall. These correlation coefficients were obtained for independent estimates and observations (validation cases). Preliminary Conclusions Using combination of MW and IR can improve both, the rainfall and snowfall estimates from an ANN model, compared to using only cloud-top IR data. The developed ANN networks do not still work well for validation case, therefore, they should be modified. Objectives Develop an artificial neural network algorithm capable to produce more accurate precipitation estimates using combination of multi-sensor infrared (from GOES) and microwave (from AMSU) cloud information. Improve IR-based rainfall retrieval algorithms using IR in conjunction with MW. Data used 1) Satellite-based cloud imagery, used as input of the ANN model: Infrared (IR) data – Cloud top brightness temperature from Channel 4 (10.7  m) from GOES 10; Microwave (MW) data – Limb corrected cloud brightness temperature from AMSU, channels frequencies 89 GHz and/or 150 GHz; 2) NEXRAD Stage-IV rainfall data is used to train (calibrate) the rainfall ANN model, and the consecutive differences in snow water equivalent (ΔSWE) data from SNOTEL stations is used to train the snowfall ANN model; 3) Maximum daily temperature data from SNOTEL stations to distinguish between snowy and rainy pixels. Study Area and Time Western United States was selected as the study area. For rainfall, the study area is located between latitudes 32º N to 42º N, and longitudes 100º W to 110º W; and for snowfall between latitudes 44º N to 49º N, and longitudes 109º W to 114º W. Total Rainfall for Summer 2004 storms selected for the study June 27-29, 2004 July 14-17, 2004 August 13-18, 2004 Acknowledgments The National Oceanic and Atmospheric Administration – Cooperative Remote Sensing Science and Technology Center (NOAA-CREST) for supporting this project. To Ralph Ferraro and Robert Kuligowski from NOAA-CICS and -NESDIS, for their valuable cooperation, comments, and providing the data. Winter season storms (Jan 01-10, 2005) were selected as the study time for snowfall, and several heavy warm season storms were selected in summer 2004 as the rainfall study time. The following images shows the cumulative rainfall amount for the selected storms from the Hydrological Data and Information System (HyDIS) website. Source: http://hydis8.eng.uci.edu/hydis-unesco/http://hydis8.eng.uci.edu/hydis-unesco/ Latitude: 32º N to 42º N Longitude: 100º W to 110º W 32º N 110º W 42º N 100º W Rainfall Study Area Snowfall Study Area Latitude: 44º N to 49º N Longitude: 109º W to 114º W 44º N 114º W 49º N 109º W Rainfall Estimates vs. Observations using only IR using 89 GHz + IR using 150 GHz + IR ΔSWE Estimates vs. Observations using only IR using 89 GHz + IR using 150 GHz + IR