Generative Modeling of Multispectral Satellite Imagery

This STTR Phase I project proposes novel deep learning models for generating realistic multi-spectral remote sensing imagery, specifically in the infrared (IR) and near-infrared (NIR) bands. The proposed system enables synthesis of semantically realistic imagery and provides parametric control of synthesizing objects-of-interest, type of terrain and land cover, time or season, weather, cloud cover, sensor parameters, and type of sensor. These parameters can be provided by the end-user or automatically generated by machine learning software modules based on the operational scenarios. The Phase I effort will focus on the development of semantic structure GAN to generate key semantic properties of the image and further extend the framework to synthesize multi-spectral imagery using the generated semantic composition based on desired criteria. The team will benefit from its vast portfolio of existing work in the areas of multispectral satellite image analysis, deep learning and convolutional neural networks, metric learning, domain adaptation, top-down and bottom-up visual attention, exploitation of remote sensing imagery, and scene understanding.