Girotto, Cristiane de Fatima (2025) Integration of Satellite Precipitation Products (SPPs) and Deep Learning (DL) for Extended Lead Time in Event-based Forecasts of Riverine Flooding Risk. Doctoral thesis, University of West London.
![]() |
PDF
Cristiane Donde Girotto - Integration of Satellite Precipitation Products (SPPs) and (April 2025).pdf - Other Restricted to Repository staff only until 31 May 2026. Available under License Creative Commons Attribution Non-commercial. Download (13MB) |
Abstract
Under current climate changing conditions, the urgent need for flood control and management calls for accurate event-based flooding forecasts with greater lead times to enable proactive strategies, which are more sustainable and less disruptive to the natural environment. However, increasing the lead time of predictions while keeping high accuracy is incredibly challenging due to the complexities of hydrological modelling and the limitations of rainfall data. These challenges are particularly significant in regions affected by severe weather originating from areas out of the range of land-based instruments. This thesis explores a groundbreaking alternative for increasing lead time in real-time predictions of riverine flooding caused by excessive
rainfall in long-lasting weather systems from ungauged regions. The study focuses on predicting rising stream water levels - a key indicator of flood risk in natural rivers -
using satellite precipitation products (SPPs), such as the Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG). While IMERG data allows
for the early detection of hazardous rainfall events in data-scarce areas, a Long Short
Term Memory (LSTM) deep learning (DL) algorithm is applied as a simple and fast alternative for real-time stream level predictions, relying solely on rainfall and stream
level data. This minimalist data-driven approach makes the model’s predictive power highly reliant on the quality and properties of the input datasets, which required the
inclusion of data pre-processing steps that are crucial for the effective application of the extensive satellite-based data. The data optimization process takes into account
the historical relationship between stream levels and precipitation data from the IMERG grid's units, or pixels, based on cross-correlation analysis. Additionally, the
effects on modelling performance of grouping the pixel data in vertical and horizontal arrangements were also examined. The method, tailored to the characteristics of the
flood prediction scenario, was fully tested and validated on predicting water level in the River Crane, near London, in the UK. The results show that the LSTM model successfully learned the historical patterns between stream levels and IMERG precipitation data collected from distances beyond 600 Km from the catchment. This ability to capture relationships over such large distances enabled real-time predictions
of stream level variations with a lead time of at least 6.5 hours using IMERG data, resulting in error rates (RMSE = 6.92 mm, MAE = 3.11 mm) comparable to those of
the model’s using rain gauge data (RMSE = 6.30 mm, MAE =2.86 mm), which offered only a 0.5 hour of lead time. Overall, the model’s performance with the NRT IMERG
data was considered 'very good' (RSR < 0.5) for real-time predictions with lead times of at least 13.5 hours. Notably, the event-based forecasts using NRT data showed
‘very good’ model performance on lead times of up to 12 hours for the October 2022 event and up to 9 hours for the October 2023 event. The method’s robustness was
confirmed through similar performances in two other catchments, located at the south and north of England. Crucially, when integrated into the real-time flood forecasting decision support platform developed by Piadeh et al. (2023a), the proposed method achieved higher accuracy and longer lead times than the original approach, which
relies on a NARX model using rain gauge data. Overall, the integration of LSTM modelling with IMERG data provided more accurate real-time predictions, offering at least a 2.5-hour increase in lead time compared to previous studies. However, the method has limitations for real-time application requiring lead times shorter than four hours, due to the latency in satellite-based rainfall estimates.
Item Type: | Thesis (Doctoral) |
---|---|
Identifier: | 10.36828/thesis/13929 |
Subjects: | Construction and engineering |
Depositing User: | Marc Forster |
Date Deposited: | 30 Jul 2025 17:04 |
Last Modified: | 30 Jul 2025 17:15 |
URI: | https://repository.uwl.ac.uk/id/eprint/13929 |
Actions (login required)
![]() |
View Item |