Deep learning models for river classification at sub-meter resolutions from multispectral and panchromatic commercial satellite imagery

Remote sensing of the Earth's surface water is critical in a wide range environmental studies, from evaluating societal impacts seasonal droughts and floods to large-scale implications climate change. Consequently, large literature exists on classification satellite imagery. Yet, previous methods have been limited by 1) spatial resolution public imagery, 2) schemes that operate at pixel level, 3) need for multiple spectral bands. We advance state-of-the-art using commercial imagery with panchromatic multispectral resolutions 30 cm 1.2 m, respectively, developing fully convolutional neural networks (FCN) can learn morphological features bodies addition their properties, FCN classify even This study focuses rivers Arctic, images Quickbird, WorldView, GeoEye satellites. Because no training data are available such high resolutions, we construct those manually. First, use RGB, NIR bands 8-band sensors. Those trained models all achieve excellent precision recall over 90% validation data, aided on-the-fly preprocessing specific In novel approach, then results model generate only require which considerably more available. Despite smaller feature space, these still 85%. provide our open-source codes parameters remote community, paves way hydrology applications vastly superior accuracies 2 orders magnitude higher than previously possible.