Dust Storm Detection Trough Moderate Resolution Imaging Spectroradiometer: a Machine Learning Problem

Dust storms are of interest to study since they are correlated to an increase in mortality rates due to respiratory illness especially in the south-western U.S. With the aim of providing better tools to the understanding of dust storms, we present models for detection of dust storms from MODIS Terra Level 1B radiances, which can be applied in near real time with 1km resolution, in contrast to those models that are based on MODIS Aerosol Optical Thickness product that are produced several hours after reception and with a 10km resolution. In this paper we present a multispectral multidimensional signal processing algorithm for detection of dust storms. The results showed that a Feed-forward Back-propagation Neural Network-based model perform better in classifying dust, and discriminating from other signatures such as clouds, smoke, etc.