Presenter: Aaron Saxton, NCSA, University of Illinois
Tuesday, May 18, 2021
Computer vision has become an essential tool for geospatial intelligence. Ever since the developments of deep learning models such as AlexNet, Inception, and ResNet50, with training sets like ImageNet, an avalanche of public sector apps have hit the market that enable identifying the faces of individuals or even allow cars to drive themselves. These apps work mostly on small images with 3 RGB channels. Imagery from satellite sensors have a unique character in that they tend to be very large images with more complicated channel configurations. Computer vision models rarely work well out of the box on these types of images. Whats more, curating training data for more specialized tasks beyond consumer grade apps still remains a major hurtle.
In this talk we will look at the anatomy of a Deep Convolutional Neural Network. Once we understand the nuts and bolts of these models we will focus on 3 aspects of model design that can help the GIS and geospatial communities improve their deep learning models as they are applied to satellite imagery. These methods are: Data augmentation, Normalization, and Hyper-parameters of convolutions. Data augmentation and normalization are important techniques to avoid overfitting. In addition to over fitting concerns, the character of satellite images can be best accounted for by adjusting hyper-parameters of a models convolutions. We will present the theory of these techniques in the context of satellite imagery to illustrate their utility.
Prerequisites: No pre-requisites.
Aaron Saxton is a Data Scientist who works in the Blue Waters project office at the National Center for Super Computing Applications (NCSA). His current interest is in machine learning, data, and migrating popular data/ML techniques to HPC environments. His career has shifted back and forth between industry and academic ventures. Previous to NCSA he was a data scientist and founding member of the agricultural data company Agrible Inc. participating in crop model development and customer facing deployment. Before that, Aaron worked at Neustar Inc, University of Kentucky, and SAIC. In the summer of 2014, shortly after joining Neustar, Aaron graduated from University of Kentucky to earn his PhD in Mathematics by studying Partial Differential Equations, Operator Theory, and the Schrodinger equation.