by Pablo Rivas Perea
Dust storms are a natural severe weather condition. A recent study in 2009 found correlation between lung diseases and... more
Dust storms are a natural severe weather condition. A recent study in 2009 found correlation between lung diseases and dust storm events. Since then, more research has been done for dust air-borne suspended particle (aerosol) analysis. However, there is paucity of formal methods in machine learning. Particularly, we study a classic and hybrid neural probabilistic approach to alleviate the lack of specialized classification methods. The maximum Likelihood Estimator (MLE) and the hybrid Probabilistic Neural Network (PNN) approaches are discussed.
The features utilized are Moderate Resolution Imaging Spectroradiometer (MODIS) thermal emissive spectral bands. We utilized four near infrared bands: B20, B29, B31, and B32. Numerical performance evaluation show that the hybrid approach (PNN) performed better than the classic (MLE). Visually, both accurately detect dust storms. The classifiers demonstrated a strong ability to find non-trivial relationships within the spectral bands. Both methods demonstrated to be soil-independent and surface-invariant detection methods. The proposed methods can be effectively utilized in understanding dust storm phenomena.
Download (.pdf) 
