613 Seminar/Meet with Speaker: Willem Marais/University of Wisconsin-Madison

Leveraging spatial textures, through machine learning, to identify aerosols and distinct cloud types from multispectral observations

Current multi-spectral methods use single pixel spectral tests and simple spatial statistical analyses to discriminate between aerosols and different cloud types from polar orbiting radiometers such as VIIRS, MODIS, AVHRR, etc. These methods misclassify extreme aerosol events as clouds (i.e. thick smoke) and distinct clouds are misclassified due similarity of reflectances and radiances. A reason for the misclassifications is that the rich spatial features of aerosols and distinct clouds are not fully characterized by simple spatial statistical analyses. Recently new machine learning methods (i.e. deep learning neural networks) have been developed to better represent spatial information of images to improve the discrimination between different image types (i.e. cars vs dogs). A new methodology has been developed to employ these new machine learning methods to improve the discriminate between aerosols and different cloud types. Specifically, 1) the NASA Worldview platform was adapted to create labeled datasets of aerosol and distinct cloud types, 2) and a pre-trained convolutional neural network (CNN) was adopted to exploit both spatial and spectral information from multi-spectral images. By harnessing CNNs with a unique labeled dataset, an improvement is demonstrated of the discrimination between aerosols and distinct cloud types from MODIS and VIIRS images compared to a method that uses spectral and simple spatial statistical tests.

Bio: Dr. Marais is a researcher at Space Science Engineering Center (SSEC) University of Wisconsin Madison. He specializes in developing novel methods to 1) denoise ground- and space-based lidar observations and 2) adapting machine learning methodologies for enhancing remote sensing methods.

613 Seminar Series Coordinators