An Image-Based Instrument for Comprehensive Weather Observations

The cloud properties are critical for climate and meteorological processes through mitigating solar radiations. This study proposed and developed a comprehensive hardware-software-algorithm system to observe cloud properties of cloud cover, cloud type, cloud bottom height and cloud move speed. The system uses contains multiple instruments of fish-eye cameras to get the whole-sky images in a very high spatial and temporal resolutions. The cloud coverage was calculated with a pixel specific deep-learning method. The cloud type was determined with an image classification deep-learning framework of DenseNet. The cloud height was derived with a multiple-angle observation method using simultaneous images captured by two stations. Finally, the cloud moving speed was calculated with two consecutive images at a station with the help of known cloud height. To validate the methods, an in-situ observation system of two coupled instruments was built up to collect real-world images. Generally, the cloud cover and cloud type respectively have an accuracy of 85% and 81% comparing against field observations of experts. The cloud bottom height was estimated with a mean error of ~250m evaluated by the data from lidar instruments. The cloud moving speed was generally correct in direction with an error level of <3m/s comparing to reanalysis data. The system provided comprehensive and reliable data of cloud in high spatial and temporal resolutions, which has promising applications in the fields of precipitation nowcast, cloud cover nowcast, and regional climate studies etc.