CNNs for Precipitation Estimation from Geostationary Satellite Imagery

The ability to detect and estimate rainfall from near-realtime satellite imagery is important in many disciplines. In this project,1 I extend recent (2016-2017) hydrometeorology work on applying deep learning to this problem by applying modern convolutional neural network (CNN) techniques. 2 After outlining the problem, I briefly discuss key prior work in both deep learning and hydrometeorology. I outline the properties of the experimental datasets. Following a discussion of the machine learning methods applied (including those of the prior work), I report my experimental results on the detection and estimation tasks. I conclude with an assessment of these results (which confirm and improve upon the prior work) and give some thoughts on future directions.