Deep-learning-based cloud segmentation and classification for weather monitoring

Traditional cloud detection algorithms for weather monitoring require radiometrically calibrated multispectral visible and infrared (IR) sensors such as those on dedicated weather satellites. In contrast, deep learning methods facilitate the use of proliferated earth observing constellations with uncalibrated sensors and fewer spectral bands for weather applications. This capability detects clouds and classifies types by recognizing spatial and spectral features using a deep neural network. In this study, we leverage existing U-Net architectures and train on several different satellite datasets. Following model development, we compare several ways to segment clouds in remote sensing images, including the number of spectral bands and separation of thin and thick clouds. The separation of thin and thick clouds is a first step in segmenting clouds by type. The capability developed in this work will facilitate the exploitation of rapidly growing data sources from the expanding market of proliferated commercial remote sensing systems.