Efficient Generation of Image Chips for Training Deep Learning Algorithms

Abstract: Training deep convolutional networks for satellite or aerial image analysis often requires a large amount of training data. For a more robust algorithm, training data need to have variations not only in the background and target, but also radiometric variations in the image such as shadowing, illumination changes, atmospheric conditions, and imaging platforms with different collection geometry. Data augmentation is a commonly used approach to generating additional training data. However, this approach is often insufficient in accounting for real world changes in lighting, location or viewpoint outside of the collection geometry. Alternatively, image simulation can be an efficient way to augment training data that incorporates all these variations, such as changing backgrounds, that may be encountered in real data. The Digital Imaging and Remote Sensing Image Image Generation (DIRSIG) model is a tool that produces synthetic imagery using a suite of physics-based radiation propagation modules. DIRSIG can simulate images taken from different sensors with variation in collection geometry, spectral response, solar elevation and angle, atmospheric models, target, and background. For our research, we selected ground vehicles as target objects and incorporated the Simulation of Urban Mobility (SUMO) model into DIRSIG to generate scenes with vehicle movement. SUMO is a multi-modal traffic simulation tool that explicitly models vehicles that move through a given road network. Using the combination of DIRSIG and SUMO, we can quickly generate hundreds of image chips, with the target at the center with different backgrounds. The simulations generated chips with vehicles and helicopters as targets, and corresponding images without targets. Using parallel computing, 120,000 training images were generated in about an hour. Some preliminary results show an improvement in the deep learning algorithm when real image training data are augmented with the simulated images.

Links to paper: