CEE Ph.D. Defense Announcement: Deep Learning Frameworks for Global Precipitation Estimation, Bias Correction and Climate Data Reconstruction from Remote Sensing

Abstract: Precipitation plays a vital role in Earth's water and energy cycles, yet accurately monitoring it remains a challenge, especially in complex terrains and remote regions. Satellite-based precipitation products (SPPs) have improved global monitoring, but their accuracy varies due to algorithmic limitations and geographic biases, particularly over complex terrains and during extreme events. This dissertation develops and evaluates deep learning approaches to improve satellite-based precipitation estimation. A bias correction framework is first applied to the PERSIANN-Dynamic Infrared Rain Rate product (PDIR-Now) over the Western U.S., demonstrating that Efficient-UNet and cGAN significantly improve accuracy by incorporating topographic data. Building on this, a new global product, PERSIANN-Unet (PUnet), is introduced, using geostationary infrared data and monthly climatology to produce high-resolution, near-real-time precipitation estimates. PUnet outperforms existing IR-based products across multiple regions. Finally, the PUnet framework is extended to reconstruct a global Climate Data Record (PUnet-CDR) from 1983 to the present, offering temporally consistent 30-minute precipitation estimates at 0.04° resolution. These advances contribute to improved monitoring of precipitation for weather forecasting, hydrologic applications and climate research.