P1.32
Multi-spectral precipitation estimation using Artificial Neural Networks
Ali Behrangi, Center for Hydrometeorology and Remote Sensing (CHRS), Irvine, CA; and K. L. Hsu, S. Sorooshian, and R. Kuligowski
Geostationary satellite sensors have been widely applied to precipitation estimation due to their capability of providing good temporal and spatial image resolution. Although most of the current geostationary based precipitation retrieval algorithms are using single or bi-spectral channels, experiments have shown that multi-spectral imagery can benefit precipitation estimation. With the advent of the next generation of geostationary satellites (e.g., GOES-R) more spectral channels with higher temporal and spatial resolution will be provided. Meanwhile, with the increased amount of information and different embedded properties, the retrieval algorithms are required to have the capability of processing multiple channels in a computationally efficient and effective manner.
A multi-spectral precipitation retrieval algorithm using an Artificial Neural Network (ANN) model has been developed. The algorithm consists of two separate stages: (1) spectral feature classification and (2) sorting classes and multi-spectral probability matching. The classification stage relies on the unsupervised self organizing feature map (SOFM) classifier. SOFM classifies image pixels, with their associated multidimensional input features, into a number of predetermined clusters. These clusters are organized into a two dimensional discrete map which preserves the topological order. Using the observed precipitation, the SOFM clusters are ranked based on their mean rain rates. Afterwards, these ranked clusters are fitted with rain observation based on the probability matching method (PMM).
The proposed algorithm was first tested using 3 months of data from the current GOES satellites with 5 spectral channels. The next step of the experiment using 12-channel spinning imager (SEVIRI) onboard EUMETDSAT's MSG satellite is ongoing. Attempts to use multi-spectral bands and their associated textural features for precipitation estimation highly increased the input space dimension. Principal component analysis (PCA) is applied to eliminate highly correlated input features. The PCA dimension compression of input spectral features also reduced the computational cost substantially. Four weeks of SEVIRI images have been evaluated and the preliminary results are encouraging. Details of the proposed algorithm, case studies, and evaluation will be presented.
Poster Session 1, Fifth GOES Users' Confererence Poster Session
Wednesday, 23 January 2008, 2:30 PM-4:00 PM, Exhibit Hall B
Previous paper Next paper