Deep Learning for Moderate Spatial Resolution Satellites: A Snow/Cloud Discrimination Example from MODIS

Abstract:

There is a burst of activity employing Deep Learning techniques to classify multispectral images. These advances have focused on satellites with high spatial resolution, where humans can produce training data of sufficient quality. Many remote sensing applications require the use of daily and sub daily products from the moderate spatial resolution, higher temporal resolution satellites (MODIS, VIIRS, GOES-R). Deep Learning techniques applied to these sensors have yet to see significant scientific interest because of the lack of training data available and the burden of manually labeling moderate spatial resolution datasets. We show that algorithms can generate the labeled data required to train from scratch a Convolutional Neural Network (CNN) for these sensors, with an example showing snow/cloud discrimination of MODIS data over the Sierra Nevada of California. We show that the high spectral fidelity and temporal resolution of MODIS enables the use of optical properties and physically based models to automatically and accurately label a nearly endless stream of data to train Deep Learning algorithms. The Snow Properties Inversion from Remote Sensing (SPIReS) algorithm is used to create a “synthetic” cloud free labeled training dataset of 20 years of daily MODIS measurements over the Sierra Nevada, which alone is a valuable product for others to use. The MOD35 cloud mask over open ocean is used to generate labeled training data of cloud objects. These labeled data are strategically used to train a CNN to produce cloud masks over snow covered terrain. Our method of generating training data, which mimics the shapes and textures and many spectral properties of the labeled classes is shown to be a viable approach to Deep Learning applications for moderate spatial resolution satellites.