Using neural network methods to estimate cloud properties from satellite imagery challenges and recent approaches

摘要: Machine learning methods, especially neural networks, have demonstrated remarkable abilities to detect and utilize spatial patterns in satellite imagery, facilitating a new generation of algorithms that estimate or forecast cloud-related properties with much higher accuracy or speed. However, we have yet to learn how to make these algorithms sufficiently robust and transparent, two properties that are crucial for responsible operational use. In this talk we hope to contribute to this learning process by highlighting some relevant challenges and promising approaches. We first highlight a few applications that use neural networks to estimate cloud properties from satellite data, including generating synthetic Multi-Radar/Multi-Sensor (MRMS) imagery from Geostationary Operational Environmental Satellites (GOES) imagery, and detecting and forecasting convection from GOES/HIMAWARI imagery. Using these …