Knowledge Discovery Framework: Deep Learning Applications for Remote Sensing

Abstract:

Earth observations (EO) have revolutionized Earth science by providing data at unprecedented temporal and spatial scales. Recent advances in the field of deep learning can be used to further enhance the impact of these data. This project, stemming from the NASA 2020 Frontier Development Laboratory summer research sprint, leverages the data scales of NASA’s EO archives to 1) develop a semantic representation encoder that can search through decades of unlabelled EO data to find phenomena of interest and 2) optimize and augment multispectral remote sensing data using imaging spectroscopy data. This encoder requires a foundation on which a robust user-friendly EO similarity search engine can be built. To this end, a self-supervised convolutional neural network was built to detect large-scale dust storms in Moderate Resolution Imaging Spectroradiometer (MODIS) images. The imaging spectroscopy subproject focuses on optimizing Sentinel-2 imagery for detecting coral reefs, as reef mapping is limited at a global scale due to the low spectral resolution of current Earth observing satellites.