Exploration of Deep Learning for Cloud Segmentation in Multispectral and Hyperspectral Satellite Imagery

Abstract

Because clouds often block much or all of the light reflected off the surface and obscure the features remote sensing seeks to reveal, typically covering 2/3 of the surface at any given time, they are one of the most significant impediments to optical remote sensing. The clouds complicate the identification of ground control points with aid and often impede the application of machine learning techniques to remote sensing data. The first step to mitigating the effect of clouds is to identify them. Several algorithms for identifying clouds have been developed, but they are inaccurate and complex enough to complicate automated data pipelines. Here, we explore how compact convolutional neural networks in the U-net architecture can identify clouds in a few examples of multispectral and hyperspectral datasets from Landsat-1 and PRISMA satellites.