Deep Learning Architectures for Improving Weather Research and Forecasting (WRF) Precipitation Estimates and Landslide Susceptibility Mapping: A Case Study of Puerto Rico

<p>Precipitation is a critical factor in the Earth’s hydrological cycle and has a profound impact on various environmental and socioeconomic systems, including management of water resources, flood forecasting, and the evaluation of natural hazards, particularly those associated with rainfall-induced landslides. For the latter, obtaining precise precipitation estimates is essential, as intense rainfall events can trigger landslides that pose significant risks to human life, infrastructure, and ecosystems, particularly in mountainous regions. Puerto Rico (PR), with its mountainous terrain, tropical climate, and vulnerability to hurricanes, provides a unique environment for studying precipitation patterns and the need for accurate rainfall estimates. The complex landscape of PR and influences from the Atlantic Ocean and Caribbean Sea lead to highly variable rainfall, highlighting the necessity of reliable precipitation data for effective hazard management and resilience planning.</p> <p>In recent years, the Weather Research and Forecasting (WRF) model has been applied in PR to support daily weather forecasting and hurricane modeling, providing critical insights into atmospheric dynamics in the region. Nonetheless, despite its numerical features, significant uncertainties persist in precipitation forecasts. These uncertainties can adversely affect the reliability of rainfall predictions, which are needed for effective hazard preparation and resource management, especially during extreme precipitation events. To address these challenges, the implementation of deep learning (DL) models utilizing high-resolution radar data offers a promising strategy for increasing the accuracy of precipitation estimations. Additionally, DL models can be employed to predict landslide susceptibility maps (LSM) that rely on forecasted precipitation, landscape characteristics, and real-time soil measurements. These maps are essential for identifying high-risk areas, enabling proactive risk management to protect lives and infrastructure.</p> <p>This dissertation aimed to develop two innovative DL models to enhance precipitation estimates generated by the WRF model and predicting landslide susceptibility in PR. The initial chapter provides a comprehensive evaluation of high-resolution precipitation products, specifically the Multi-Radar Multi-Sensor and Stage IV datasets. This evaluation employs a rain gauge network established by the United States Geological Survey to assess the accuracy of precipitation products across various rainfall categories in PR. The findings yield valuable insights into the reliability of radar-based products, establishing a foundation for improving WRF precipitation estimates. The second chapter outlines a novel DL architecture that incorporates attention gates to improve feature extraction and downscale precipitation estimates. This model relied on WRF output variables and high-resolution radar products as ground truth, supported by a novel customized percentile loss function to ensure improvements during extreme precipitation events. Finally, the third chapter focusses on the application of new updated WRF precipitation data in generating LSM through an optimized DL model. This model integrates a landslide inventory that was compiled after Hurricane Maria in 2017, high-resolution landscape data, and previous-day soil moisture and saturation data obtained from GOES-PRWEB, alongside the corrected forecasted precipitation data. The model effectively predicts LSM throughout PR prior to extreme precipitation events, providing a critical tool for proactive risk management by combining landscape characteristics, real-time soil data, and new improved WRF precipitation data.</p>

Name thesis:degree_name

Doctor of Philosophy (Ph.D.)

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Civil Engineering

Year dc:date.available

2025

Author dc:creator

• Mejia Manrique, Said A

Contributors dc:contributor

• Reza Khanbilvardi

• Deep Learning Models
• Weather Research and Forecasting (WRF) Model
• Radar- Based Precipitaion Products
• Lanslide Susceptibility Mapping
• Civil Engineering

Repository record dc:identifier

https://academicworks.cuny.edu/cc_etds_theses/1286

OAI identifier oai:identifier

oai:academicworks.cuny.edu:cc_etds_theses-2283