Blog

Dust Source Identification

In this study, we take a radically different approach to the previous studies by using machine learning to objectively provide a multi-variate and non-linear classification of surface types using multi-spectral satellite data. Our ultimate goal is to identify all the surface locations on the planet that are dust sources. To do this we use a Self Organizing Map (SOM) Neural Network to classify all the land surface locations into a set of 1,000 categories, a small subset of these categories will be encompassing regions that are dust sources. There are a variety of types of dust sources (e.g. dry river beds, agricultural sources, etc.) that we would like to delineate.

To achieve this comprehensive classification we consider the conditions present throughout an entire year of the 0.05 degree MCD43C3 data product. The movies to the right show how the bihemispherical reflectance varies through the year for the 7 bands provided in the MCD43C3 MODIS product. This is a massive dataset, and the time required to perform the SOM classification increases with the number of data records. We therefore first restricted our attention to those broad MODIS surface types that are likely to include all dust sources, namely: barren or sparsely vegetated surfaces, croplands, grasslands, and open and closed shrublands. These are MODIS surface types 16, 12, 10, 7 & 6 respectively. For each of these surface types we then construct an input vector that contains 7 values, namely the 7 bands provided in the MCD43C3 MODIS bihemispherical reflectance. Once we have provided the SOM with these inputs it returns to us the 1,000 class definitions. To limit the classes we compare them to the Naval Research Laboratory (NRL) high-resolution dust source databases (DSD) for South West Asia and East Asia. Those classes that overlap sources in the NRL DSDs are weighted and ranked. With this approach we found good agreement with dust sources previously identified in North Africa and as compared with the NRLs dust enhancement satellite products derived from MODIS and Meteosat Second Generation data.

An interesting result using this first guess approach is the exclusion of classes that overlaid sources in the Sahel as seen in the dust enhancement imagery. Such a result implies dust sources in the Sahel have a different reflectance to surface type classification than those sources in South West Asia and East Asia. Further discussion will be given on dust source areas in North and South America and Australia including their classifications with respect to reflectance and surface type.

Dust Source Identification