A Classic and Neural Probabilistic Approach to the Dust Storm Detection Problem

This paper address the problem of dust storm detection based on multispectral image analysis from a probabilistic point of view. Two classifiers are designed, one based on classic probability theory and other based on a probabilistic computational intelligence approach. The first classifier is designed under the Maximum Likelihood Estimation (MLE) model, and the second with a Probabilistic Neural Network (PNN) model. The data set used in this work consists of MODIS instrument at the NASAs Terra satellite data, generating 75 millions of samples used in the design and validation of the classifiers. Findings indicated that the PNN presents a better classification performance than the MLE classifier. The proposed models are suitable for near real-time applications, and provide with an output at a resolution of 1km, which is an improvement over the methods based on the MODIS AOT product which has a 10km resolution.