Generating Synthetic Training Data for Satellite Imagery Applications

Abstract:

Supervised machine learning (ML) models rely on high quality training data to learn the relationship between input and target variables. In many remote sensing applications, such as agricultural monitoring, data collected on the ground (a.k.a. ground reference) are matched with corresponding satellite imagery to form a training dataset. However, ground reference data collection is an extensive and expensive effort, and extremely scarce in remote and dangerous areas that would most benefit from remote sensing. In this study, we use Generative Adversarial Networks (GAN) to generate synthetic training data from multispectral satellite imagery.

We developed two techniques applied to Sentinel-2 satellite observations: a) generating unlabeled synthetic multispectral images at the native resolution of Sentinel-2 data, and b) conditioning the synthetic image generation on its land cover label. Applying both techniques to data in Midwest US and Western Cape South Africa we show the success of using a GAN architecture to generate synthetic labeled satellite imagery. We will use the synthetic data in combination with labeled training data to show the impact on deep learning based classification model’s accuracy for problems that use satellite imagery. In the next phase, we plan to use the same approach and generate synthetic training data with labels to improve the performance of semantic segmentation models.