A Discriminative Modeling Framework for Mining of Spatio-Temporal Data in Remote Sensing

Supported by the National Science Foundation (NSF III-Small) IIS-1117433

Duration: 3 years

Investigators

Slobodan Vucetic

Department of Computer and Information Sciences

Temple University

303 Wachman Hall

1805 N. Broad Street

Philadelphia PA 19122

 

Zoran Obradovic

Department of Computer and Information Sciences

Temple University

303 Wachman Hall

1805 N. Broad Street

Philadelphia PA 19122

Project Summary

Understanding physical and chemical processes of Earth’s atmosphere, land, and ocean is a scientific challenge of utmost importance to society. This is reflected by significant investments in a large number of satellite and ground-based sensors dedicated to Earth observation. Through a process called retrieval, the observations by these sensors are used to estimate important geophysical properties such as temperature, clouds, aerosols, greenhouse gases, snow and ice, and biological productivity. The retrieved values are then used as input data in various scientific studies such as climate modeling, weather forecasting, air quality monitoring, and disease management. Maximizing the quality of retrievals, where the quality is measured by retrieval accuracy and ability to estimate retrieval uncertainty, is therefore critical for success of a wide range of scientific studies. The objective of this project is to best utilize large quantities of multi-source observations from satellite and ground-based instruments for Earth observation having different capabilities regarding coverage, resolution, and quality. The traditional techniques from spatial statistics and data fusion are not satisfactory, due either to computational constraints, difficulties in modeling and parameter estimation, or inability to provide uncertainty estimates.

     This project will develop a novel discriminative modeling framework for fusion of multi-sensor remote sensing data based on the Gaussian conditional random field model. It will be designed to become highly flexible, robust, and computationally efficient, thus enabling its use on large spatio-temporal data sets. It will allow learning from and predicting in the presence of a mixture of labeled and unlabeled data with partially observable attributes, utilizing large quantities of unlabeled data, and dealing with sampling bias. Following the design phase, the framework will be transformed into a flexible and powerful software tool that will be easy to use and understand by practitioners. The proposed data fusion approach will be applied and evaluated on the problems of atmospheric aerosol retrieval and surface level pollution estimation, which are the high-impact challenges of climate research and environmental science.

     The broader impact of this research will be in methodological advancements in spatio-temporal data mining and geostatistics and advancement of the state of the art in Earth remote sensing. Considering the scientific importance and effort invested in Earth observation, maximizing the utility of multi-source observations is a top-priority scientific challenge. Through improvements in aerosol retrieval from multi-sensor data, this project will enable improved characterization of the effects of aerosols on the Earth’s radiation budget and climate. Beyond aerosol retrieval, the developed framework will be directly applicable to retrievals of many other atmospheric, land, and ocean properties. In addition, the new approach will be highly applicable to other scientific and engineering domains where a large volume of data is collected by multiple instruments over time and space to learn the properties of an underlying complex phenomenon. The activities in this project will be integrated with education and aimed toward a broad participation of students ranging from doctoral to K-12 level and increasing diversity by using the already established channels at Temple University for involving underrepresented students.