On the enhancement of infrared satellite precipitation estimates using genetic algorithm filter-based feature selection

A methodology to enhance a satellite infrared - based high resolution rainfall retrieval algorithm is developed by intelligently selecting features based on a filter model. Our methodology for satellite-based rainfall estimation is similar to the PERSIANN-CCS approach. However, our algorithms are enriched by applying a filterbased feature selection using generic algorithm. The objective of using feature selection is to find the optimal set of features by removing the redundant and irrelevant features. Since we use unsupervised cloud classification technique, Self Organizing Map (SOM), an unsupervised feature selection method, is used. In our approach, first the redundant features are removed by using a feature similarity-based filter and then using Entropy Index along with genetic algorithm searching, the irrelevant features are eliminated. The result shows that using feature selection process can improve Rain/No Rain detection about 10 % at some threshold values and also decreases the RMSE about 2mm.