Compared to ground-based precipitation measurements, satellite-based precipitation estimates have the advantage of global coverage and high spatiotemporal resolutions. However, the accuracy of satellite-based precipitation observations is still insufficient to serve many weather, climate, and hydrologic applications. In the development of a satellite-based precipitation product, the two most important aspects are the provision of sufficient precipitation-related information in the selected satellite data and the use of the proper methodologies to extract such information and link it to precipitation estimates.
In this dissertation, a state-of-the-art deep learning framework for precipitation estimation using bispectral satellite information, Infrared (IR) and water vapor (WV) channels, is developed. I explore the effectiveness of deep learning techniques in extracting useful features from the satellite information and demonstrate the value of incorporating multiple satellite channels.
Specifically, I first provide a bias reduction model for satellite-based precipitation products using deep learning approaches to demonstrate their capability of extracting additional useful information from the satellite data. I then design a two-stage framework for precipitation estimation from bispectral information, consisting of an initial rain/no-rain (R/NR) binary classification, followed by a second stage estimating the non-zero precipitation amount. In the first stage, the model aims to eliminate the large fraction of NR pixels and to precisely delineate precipitation regions. In the second stage, the model aims to estimate the point-wise precipitation amount accurately while preserving its heavy-tailed distribution. Stacked denoising auto-encoders (SDAEs), a commonly used deep learning method, are applied in both stages.
The operational product, Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks Cloud Classification System (PERSIANN-CCS), serves as a baseline model throughout this dissertation. I evaluate performance along a number of common performance measures, including both R/NR and real-valued precipitation accuracy. Case studies focusing on the model’s performance for specific events are also included. The experiments show that our proposed two-stage model outperforms original PERSIANN-CCS in different verification periods over the central United States and in large-scale application. Therefore, the two-stage deep learning framework has the potential to serve as a more accurate and more reliable satellite-based precipitation estimation algorithm.