Accurate and timely precipitation estimates are critical for many hydrological applications, including monitoring and forecasting natural disasters, developing water resources management and planning strategies, as well as conducting climatological studies. Despite having high-resolution satellite information, precipitation estimation from remotely sensed data still suffers from methodological limitations. State-of-the-art deep learning algorithms, renowned for their skill in learning accurate patterns within large and complex data sets, appear well-suited to the task of precipitation estimation. By leveraging the deep learning algorithms with an ample amount of high-resolution satellite datasets, I introduce one historical and one near-real time precipitation estimation dataset: 1) The near-real time dataset called Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Convolutional Neural Networks (PERSIANN-CNN). This developed near- real time dataset provides precipitation estimates at 0.08-degree spatial and an hourly temporal resolution. 2) The historical precipitation estimation dataset named Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Cloud Classification System-Climate Data Record (PERSIANN-CCS-CDR). This dataset offers precipitation estimates at 0.04° spatial and 3-hourly temporal resolutions from 1983 to present over the globe.Chapter 2 and 3 investigate the application of two different deep neural networks algorithms for improving the near-real time precipitation estimation. Most near real-time precipitation retrieval algorithms utilize infrared (IR) information as their input due to its fine spatiotemporal resolution and near-instantaneous availability. Their sole reliance on IR information is problematic. Indeed, it limits their ability to learn different mechanisms of precipitation during training, resulting in less accurate estimates. However, recent advances in the field of machine learning offer attractive opportunities to improve the precipitation retrieval algorithms. Chapters 2 investigates the effectiveness of adding Water Vapor (WV) channels from geostationary satellites to IR information and the application of convolutional neural networks for improving the accuracy of near real-time precipitation algorithms. Chapter 3 further improves the model introduced in Chapter 2 by adding geographical information (i.e., latitude and longitude) to IR information and utilizing a U-Net-based convolutional neural network. The developed dataset called PERSIANN-CNN, which provides near-real time precipitation estimates at the spatial resolution of 0.08-degree and at an hourly time scale over the CONUS. Results demonstrate that the proposed model (PERSIANN-CNN) provides more accurate rainfall estimates compared to the current operational near-real time precipitation datasets at various temporal and spatial scales. Chapter 4 and 5 introduce a new historical precipitation estimation dataset through introducing a new framework for bias correcting and downscaling the precipitation estimates. Accurate, long-term, global precipitation estimates, especially for heavy precipitation rates, at fine spatial and temporal resolutions is vital for a wide variety of climatological studies. Most of the available operational precipitation estimation datasets provide either high spatial resolution with short-term duration estimates or lower spatial resolution with long-term duration estimates. Furthermore, previous research has stressed that the available satellite-based precipitation products show poor performance for capturing extreme events at high temporal resolution. Therefore, there is a need for a precipitation product that reliably detects heavy precipitation rates with fine spatiotemporal resolution and a longer period of record. Chapters 4 and 5 introducing a new historical dataset that address these limitations by introducing a bias correcting and a downscaling framework, respectively. The developed dataset (called PERSIANN-CCS-CDR) provides precipitation estimates at 0.04° spatial and 3-hourly temporal resolutions from 1983 to present over the global domain of 60°S to 60°N. PERSIANN-CCS-CDR shows improved performance compared to PERSIANN-CDR in representing the spatiotemporal resolution, magnitude, and spatial distribution patterns of precipitation, especially for extreme events. In chapter 6, the applications of the developed datasets as well as PERSIANN family datasets are assessed to investigate the spatiotemporal variations in heavy precipitation events that occurred in early spring (March 21st to April 20th) over Iran. The Results show that PERSIANN family datasets are an attractive dataset for detecting the near-real time and historical precipitation estimates.

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.

Accurate precipitation estimation at fine spatial and temporal scale is crucial for climatological studies. The Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Climate Data Record (PERSIANN-CDR) is a well-known estimation product and is bias-corrected using the Global Precipitation Climatology Project (GPCP), which has been recently updated to version 2.3. In this study, we compare the PERSIANN-CDR dataset that is bias-corrected with GPCP V2.3 (called PERSIANN-CDR V2.3) with the previous version, which was bias-corrected by GPCP V2.2 (PERSIANN-CDR V2.2), at monthly and daily scales. First, we discuss the changes between the two versions of PERSIANN-CDR using Mean Absolute Difference (MAD) and Relative Mean Absolute Difference (MARD) at the monthly scales over the globe. The results show noticeable differences between PERSIANN-CDR V2.3 & V2.2 over the ocean for latitudes from 40 to 60 after 2003. The changes are also significant over the land area from 2009 onward. Second, we evaluate the improvements in the new version of PERSIANN-CDR (V2.3) with respect to a gauged-based reference, data from Climate Prediction Center (CPC), at monthly and daily scales over all globe land areas and again over CONUS. Over the globe, the estimation of PERSIANN-CDR V2.3 is more accurate than PERSIANN-CDR V2.2, especially over CONUS and Australia. Over CONUS, Root Mean Square Error (RMSE) has decreased by 4.3 % and Correlation Coefficient (CC) has improved by 3.8 % compared to PERSIANN-CDR V2.2. The results emphasize that PERSIANN-CDR V2.3 has significantly improved in performance owing to refinement and input data from GPCP beginning from 2003.

Space-based observations, emerged in the hydrology field in the last two decades, play a fundamental role in providing alternative information of hydrologic variables besides gauge measurements, especially in data scarce regions. Among all satellite products, products derived from Moderate Resolution Imaging Spectroradiometer (MODIS) Satellite are popular due to the satellite's rapid re-visit time and adequate spatial resolutions. However, cloud obscuration limits the usage of products derived from MODIS because clouds block satellites from capturing the ground state of the earth surface. This dissertation aims to (1) recover two cloud-free MODIS datasets of snow and flood using a 3-D interpolation technique, namely, Variational Interpolation (VI) and (2) demonstrate their usefulness for hydrologic applications.

In the first part of this dissertation, the computational stability of the existing VI method is improved, then, we apply the algorithm to produce a cloud-free snow dataset for CONUS from 2000 to 2017. Moreover, by taking into consideration specific assumptions about the water body characteristic, we implement VI algorithm to remove clouds from MODIS flood maps. Promising results from a validation period over the Mississippi River are presented. We also couple the elevation information to derive the cloud-free MODIS water depth maps from the MODIS water extent maps. Water level maps are important for hydrological studies and can also act as references when the future Surface Water and Ocean Topography (SWOT) direct observations of water elevation are available in 2021.

In the second part of this dissertation, we use the resulting cloud-free MODIS flood and water depth maps to improve a hydrological model by reducing model errors via calibration and data assimilation. The calibrated output inundation maps accurately reflect flood events for the Upper Mississippi River Basin in 2013 and 2014. Also, the downstream discharge via the data assimilation scheme can correctly predict flood events during the same validation period. The results indicate that the framework can be further used to monitor and forecast floods.

Floods are among the most devastating natural disasters which affect millions of people worldwide. Forecasting floods to provide warnings to the public in a timely manner is crucially important, however, this is a very challenging task. HiResFlood-UCI was developed by coupling the National Weather Service's (NWS) distributed hydrologic model (HL-RDHM) with the hydraulic model BreZo (developed by Sanders and Begnudelli) in order to estimate localized flood depths and velocities. A semi-automated technique of efficient unstructured mesh generation for BreZo was developed. HiResFlood-UCI was implemented for the ELDO2 catchment in Oklahoma. Using synthetic precipitation input, the model was tested for various components including HL-RDHM parameters (a priori versus calibrated), channel and floodplain Manning n values, DEM resolution (10m versus 30m), and computation mesh resolution (10m+ versus 30m+). Simulations show that HiResFlood-UCI produces reasonable results with the a priori parameters from NWS. Sensitivities to hydraulic model resistance parameters, mesh resolution, and DEM resolution are also identified, pointing to the importance of model calibration and validation for accurate prediction of localized flood intensities. HiResFlood-UCI performance was examined using six measured precipitation events as model input for validation of the streamflow at the outlet and an interior point. Validation builds confidence in model predictions of river discharge, flood extent and localized velocities, which are fundamental to reliable flood warning.

HiResFlood-UCI was implemented for flood forecasting in the Cedar River Basin using real-time remote sensing precipitation PERSIANN-CCS data. The model was evaluated for the historical 2008 Iowa flood. The results show HiResFlood-UCI with real-time PERSIANN-CCS was able to capture the observed hydrographs and reasonably match the USDA's AWiFS 56m resolution flood imagery over the most impacted area in the extended Cedar Rapids region. This is promising for a global high resolution flood warning system pairing HiResFlood-UCI with PERSIANN-CCS in the near future.