Award: IIS-0612149

Title: Data Mining Support for Retrieval and Analysis of Geophysical Parameters

Investigators: Zoran Obradovic¹ (PI), Slobodan Vucetic¹ (Co-PI), Zhanqing Li² (Co-PI), Bo Han¹ (Ph.D. student), Vladan Radosavljevic¹ (Ph.D. student)

¹Center for Information Science and Technology, Temple University, Philadelphia, PA 19122

²Dept. of Meteorology, University of Maryland,  College Park, MD 20742

Motivation and objective: Aerosols, small particles emanating from natural and man-made sources have been recognized as one of the major factors in ongoing climate studies. By reflecting and absorbing solar radiation, aerosols have direct influence in both cooling the surface and warming the atmosphere and indirect influence on cloud properties and precipitation. This project aims at facilitating estimation (retrieval) of aerosol optical thickness (AOT) at a global scale through development of data mining methods to improve existing single-sensor satellite-based retrieval algorithms and to allow higher-quality retrieval using multiple satellite and ground -based sensors.

Approach: In the first 5 months of the project, we analyzed retrieval error of operational aerosol retrieval methods and developed statistical retrieval method that makes use of aerosol observations collected by AERONET ground-based radiometers at about 180 sites worldwide (locations of US sites showed at Fig. 1a) and reflected solar radiation measured by multiple spectral instruments (MODIS and MISR) aboard AQUA and TERRA satellites (Fig. 1b). The appeal of using both types of data is that ground-based measurements are of high quality, while satellite-based observations are of high spatial resolution. The main challenge addressed in out research is how to explore spatial-temporal properties of such heterogeneous and multi-resolution data.

                                           a)                                                                    b)

Figure 1.a) Location of ground-based AERONET sites in US and b) reflected solar radiation as observed by satellite and AERONET instruments.

New results: Our method for discovering sources of retrieval errors consists of: 1) learning relationships between satellite observations and AOT by neural networks; 2) identifying conditions when the neural network is more accurate than operational retrieval, as in such situations accuracy of the later can be further improved.  The new method was tested on spatially and temporally collocated MODIS Collection 004 and AERONET data over continental US between 2002 - 2004. In our experiments, significantly more accurate retrievals were obtained by neural networks. Our analysis resulted in a partial understanding of conditions where operational MODIS Collection 004 algorithm can be improved. This includes cases when MODIS AOT retrieval has high values, when Angstrom exponent is large, over areas contaminated with clouds, and over desert areas.

Recently, new MODIS Collection 005 retrieval algorithm has been developed. Our comparison over continental US during the first six months in year 2005 suggests that Collection 005 retrievals are much more accurate than Collection 004, but neural network-based retrievals are still more accurate (Fig. 2). These results indicate that, despite the recent advances, the quality of operational retrievals can be further improved and in the following months we will develop a set of recommendations that should lead to further improvement of MODIS operational retrievals.

                      a)                                            b)                                            c)

Figure 2. Comparative results of a) MODIS Collection 004 b) MODIS Collection 005 and c) Neural Network-based AOT retrievals over continental US between January-June 2005.