UC Irvine

Clouds and precipitation are key components of the Earth’s hydrological cycle, yet there is a lack of deep understanding of their physics and dynamics. Reliable cloud and precipitation measurements are useful for understanding the current state of global water resources and determining the characteristics of these processes over time and space. Over the past few decades, different research groups have been attempting to develop an optimal algorithm for deriving near-real-time precipitation at a global scale; yet, due largely to the tradeoff between passive and active remotely sensed information, the task remains elusive. Rapid developments in satellite observational technologies, along with advancements in machine learning techniques and computational power, offer opportunities to integrate massive amounts of real-time observations to characterize spatiotemporal structures of clouds and their associated precipitation.In the first half of this dissertation, the performance of well-known operational precipitation algorithms is assessed during Atmospheric River events, classified into brightband and nonbrightband systems. The inter-comparison of high spatiotemporal resolution surface precipitation products has shown the significant discrepancies and limitations of satellite algorithms to retrieve shallow precipitation systems, especially over complex terrains (Chapter 2). Next, this dissertation introduces a Deep Neural Network algorithm that integrates the vertical properties of clouds, derived from the unique NASA CloudSat Cloud Profiling Radar, with multispectral images derived from the new generation of geostationary satellites used for near-real-time cloud-type classification—including cirrus, altostratus, altocumulus, stratus, stratocumulus, cumulus, nimbostratus, and deep convective clouds (Chapter 3). Our analysis suggests that the model, named Deep Neural Network Cloud-Type Classification (DeepCTC), provides supplementary insights into the variability of cloud types to diagnose the weaknesses and strengths of near-real-time geostationary-based precipitation retrievals. In the latter half of this paper, an “end-to-end” deep learning architecture is developed that fuses infrared band from geostationary platforms with low-Earth-orbit passive microwave measurements to retrieve surface precipitation rates in 4km spatial resolution (Chapter 4). This model, Deep Neural Network High Spatiotemporal Resolution Precipitation Estimation (Deep-STEP), shows a promising opportunity for capturing fine-scale precipitation patterns over complex surface types such as coastal regions. However, the model did not perform well in detecting surface precipitation detection during warm months. Chapter 5 consequently presents a refined model, which can better estimate summertime surface precipitation, and explores this model’s potential by adding a wide range of multispectral information covering visible, near infrared, and infrared portions of the electromagnetic spectrum (as well as auxiliary variables from Global Forecast System) into the model. The results highlight the model’s ability to extract fine-scale spatial patterns of surface precipitation over convective and stratiform events. In addition, our analysis demonstrates significant improvements in capturing the occurrence and amount of precipitation over orographic and coastline regions when compared to the operational algorithms. We anticipate our investigation to be a starting point for more sophisticated and efficient precipitation retrieval systems in terms of accuracy, fine spatial pattern detection skills, and computational costs.