Feature Extraction for Extreme Precipitation Events using Deep-Networks

This work presents a study about extreme rainfall events in the Colombia southwestern between 1983 and 2019 using satellite information from CHIRPS. The information allows getting the standardized precipitation index (SPI) for four time-scales: monthly, trimestral, semestral, and annual, which is necessary to understand how spatiotemporally is wet or dry a place. Due to a large amount of data, we used a dimensional reduction approach based on neural networks knows as Non-Linear PCA to get the principal components for each scale and make five clustering procedures to identify regions with similarities. We choose the number of clusters from different clustering metrics. Although for SPI1 and SPI3, the results were inconsistent, the results for SPI6 y SPI12 were quite good. The SPI6 got two regions: West and East, while the SPI12 got five regions: Pacific South, Pacific North, Andean, Andean foothills, and Amazon Regions. The findings also show differences in frequency, duration, and intensity of extreme events. Thus, we conclude that for SPI6, the East region is dryer than the West one, and for SPI12, the Andean region is the driest while the Pacific South is the wettest.