On The Use Of Image Processing And Pattern Recognition Tools To Enhance High Resolution Satellite Precipitation Estimation Based On Cloud Classification

Satellite precipitation estimation at high spatial and temporal resolutions is beneficial for research and applications in the areas of weather, flood forecasting, hydrology, and agriculture. In this research, image processing and pattern recognition tools are incorporated into the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks Cloud Classification System (PERSIANN-CCS) methodology to enhance satellite precipitation and rainfall estimation. The enhanced algorithm incorporates five main steps to derive precipitation estimates: 1) segmenting the satellite infrared cloud images into patches; 2) extracting features from the segmented cloud patches; 3) feature selection or dimensionality reduction; 4) categorizing the cloud patches into separate groups; and 5) obtaining a relationship between the brightness temperature of cloud patches and the rain- rate (T-R) for every cluster. In this study, in addition to the features utilized for cloud patch classification, wavelet and lightning features are also extracted. The lightning feature is defined as the number of flashes occurring within 15 minutes of the nominal IR image scan. Both feature selection and dimensionality reduction techniques are examined to reduce the dimensionality as well as diminish the effects of the redundant and irrelevant features. The feature selection technique includes a Feature Similarity Selection (FSS) method and a Filter-Based Feature Selection using Genetic Algorithm (FFSGA). The Entropy Index (EI) fitness function is used to evaluate the feature subsets. Furthermore, Independent Component Analysis (ICA) was examined and compared to other linear and nonlinear unsupervised dimensionality reduction techniques to reduce the dimensionality and increase the estimation performance. In addition to a Self Organizing Map (SOM) neural network, the link-based cluster ensemble method is also examined in this research. In the final step, the Median Merging (MM) and Selected Curve Fitting (SCF) techniques are incorporated. After applying a Probability Matching Method (PMM) to each single patch and obtaining the T-R for each patch, a Median Merging technique which computes the median rain-rate for a given temperature is applied. A Selected Curve Fitting (SCF) procedure is also used to obtain the T-R for each cluster. The results show that the enhanced algorithm incorporating the above techniques improves precipitation estimation