Lucy Marshall
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Lucy Marshall has a total of 237 co-authors. Here are the top co-authors by number of shared publications:
- Ashish Sharma
School of Civil and Environmental Engineering University of New South Wales Sydney NSW Australia
49 shared publications - David John Nott
Department of Statistics and Applied Probability, National University of Singapore, Singapore
12 shared publications - Fiona Johnson
Water Research Centre, School of Civil and Environmental Engineering, UNSW Sydney, Australia
12 shared publications - Tyler Smith
Department of Civil and Environmental Engineering; Clarkson University; Potsdam New York USA
10 shared publications - Clare Stephens
Hawkesbury Institute for the Environment, Western Sydney University, Richmond, NSW, Australia
9 shared publications - Kristen D. Splinter
Water Research Laboratory School of Civil and Environmental Engineering UNSW Sydney Sydney NSW Australia
7 shared publications
Latest Publications
- Open Access
- 5 September 2026
- Published by Elsevier in Environmental Modelling & Software
- Vol. 206
UrbanFloodBench: Bridging AI and hydrology through benchmarking of coupled 1D–2D urban flood surrogate models
- 5 September 2026
- Published by Elsevier in Environmental Modelling & Software
- Vol. 206
Open AccessUrbanFloodBench: Bridging AI and hydrology through benchmarking of coupled 1D–2D urban flood surrogate models
- Open Access
- 4 August 2026
- Published by Elsevier in Environmental Modelling & Software
- Vol. 205
Exploring graph neural networks for flood modeling: Challenges, opportunities, and future prospects
- 4 August 2026
- Published by Elsevier in Environmental Modelling & Software
- Vol. 205
Open AccessExploring graph neural networks for flood modeling: Challenges, opportunities, and future prospects
- Open Access
- 8 June 2026
- Published by Elsevier in SSRN Electronic Journal
UrbanFloodBench: Bridging AI and Hydrology through Benchmarking of Coupled 1D–2D Urban Flood Surrogate Models
Physics-based hydrodynamic models are the standard approach for urban flood simulation but remain computationally prohibitive for real-time forecasting. Although machine learning surrogate models offer a promising alternative, most existing approaches rely on private or simplified datasets, limiting reproducibility and rigorous comparison, particularly for coupled 1D--2D flood systems. To address this gap, we introduce \textit{UrbanFloodBench}, the first open benchmark for coupled 1D--2D urban flood surrogate modelling. UrbanFloodBench combines a multi-catchment dataset, a variance-aware evaluation framework, and a reproducible large-scale AI Kaggle challenge. The benchmark involves 264 teams and 3,799 submissions, producing the first large-scale empirical comparison of AI methods for coupled urban flood emulation. Results show that multiple modelling approaches can achieve strong performance, while the 1D drainage component remains the primary unresolved challenge. UrbanFloodBench establishes a reproducible foundation for advancing AI-based surrogate modelling for urban flood prediction and enabling consistent future evaluation of data-driven flood forecasting methodologies.
- 8 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessUrbanFloodBench: Bridging AI and Hydrology through Benchmarking of Coupled 1D–2D Urban Flood Surrogate Models
Physics-based hydrodynamic models are the standard approach for urban flood simulation but remain computationally prohibitive for real-time forecasting. Although machine learning surrogate models offer a promising alternative, most existing approaches rely on private or simplified datasets, limiting reproducibility and rigorous comparison, particularly for coupled 1D--2D flood systems. To address this gap, we introduce \textit{UrbanFloodBench}, the first open benchmark for coupled 1D--2D urban flood surrogate modelling. UrbanFloodBench combines a multi-catchment dataset, a variance-aware evaluation framework, and a reproducible large-scale AI Kaggle challenge. The benchmark involves 264 teams and 3,799 submissions, producing the first large-scale empirical comparison of AI methods for coupled urban flood emulation. Results show that multiple modelling approaches can achieve strong performance, while the 1D drainage component remains the primary unresolved challenge. UrbanFloodBench establishes a reproducible foundation for advancing AI-based surrogate modelling for urban flood prediction and enabling consistent future evaluation of data-driven flood forecasting methodologies.
- Open Access
- 9 April 2026
- Published by Elsevier in SSRN Electronic Journal
Hydrologically constrained genetic programming for interpretable rainfall--runoff model discovery: A process-informed machine learning approach
Advances in machine learning have improved rainfall–runoff prediction, but integrating these capabilities with interpretable, process-based model structures remains an open challenge. This study seeks a middle ground between conceptual hydrological modelling and modern data-driven prediction by automating model structure discovery while retaining interpretability and hydrological consistency. We introduce HydroXGP (Hydrological Model eXplorer using Genetic Programming) to reduce reliance on legacy model choices and subjective a priori structural assumptions by learning model structure directly from data during optimization. The framework evolves models from a modular, process-based component library adapted from MARRMoT (Knoben et al., 2019), thereby making the search process-informed through reusable flux and routing components and hydrologically constrained by evaluation under mass balance. GP jointly optimizes model structure and parameters, while each discovered model remains explicit and interpretable. Evaluation on the CAMELS-US dataset showed that a proof-of-concept test in a snow-dominated basin achieved strong performance over the out-of-sample test period (NSE=0.849), and, when transferred to a hydrologically similar basin, the learned structure retained useful predictive skill (NSE=0.724). Across 18 basins (one representative basin per U.S. hydrologic region), median NSE reached 0.702, generally outperforming the mean performance of an ensemble of conceptual benchmark models while remaining competitive with regional Long Short-Term Memory (LSTM) baselines. The findings support a modelling perspective in which structural hypotheses can be revisited and adapted as new data, diagnostics, or environmental signals become available, enabling basin-specific configurations that evolve in response to changing information rather than relying on a single fixed template.
- 9 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHydrologically constrained genetic programming for interpretable rainfall--runoff model discovery: A process-informed machine learning approach
Advances in machine learning have improved rainfall–runoff prediction, but integrating these capabilities with interpretable, process-based model structures remains an open challenge. This study seeks a middle ground between conceptual hydrological modelling and modern data-driven prediction by automating model structure discovery while retaining interpretability and hydrological consistency. We introduce HydroXGP (Hydrological Model eXplorer using Genetic Programming) to reduce reliance on legacy model choices and subjective a priori structural assumptions by learning model structure directly from data during optimization. The framework evolves models from a modular, process-based component library adapted from MARRMoT (Knoben et al., 2019), thereby making the search process-informed through reusable flux and routing components and hydrologically constrained by evaluation under mass balance. GP jointly optimizes model structure and parameters, while each discovered model remains explicit and interpretable. Evaluation on the CAMELS-US dataset showed that a proof-of-concept test in a snow-dominated basin achieved strong performance over the out-of-sample test period (NSE=0.849), and, when transferred to a hydrologically similar basin, the learned structure retained useful predictive skill (NSE=0.724). Across 18 basins (one representative basin per U.S. hydrologic region), median NSE reached 0.702, generally outperforming the mean performance of an ensemble of conceptual benchmark models while remaining competitive with regional Long Short-Term Memory (LSTM) baselines. The findings support a modelling perspective in which structural hypotheses can be revisited and adapted as new data, diagnostics, or environmental signals become available, enabling basin-specific configurations that evolve in response to changing information rather than relying on a single fixed template.
- Open Access
- 13 February 2026
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 3 (1)
A Generalized Data‐Driven Shoreline Model at the Regional Scale
Shoreline models underpin long‐term coastal planning and hazard mitigation but are often limited by site‐specific calibration and simplified process representations. While machine learning (ML) approaches have recently matched or exceeded the skill of traditional models in deterministic shoreline forecasting, they have yet to demonstrate generalization across diverse coastal environments. In this study, we introduce a generalized, data‐driven regional‐scale shoreline model trained across 2,000 km of the southeast Australian coastline. The model adapts the Temporal Fusion Transformer (TFT) architecture, combining a sequence‐to‐sequence Long Short‐Term Memory (LSTM) module with masked multi‐head self‐attention to learn both short‐ and long‐range temporal dependencies. Static covariates unique to each transect are integrated via gating mechanisms that condition the learned relationship between wave action and shoreline response, enabling outputs to adapt to local site characteristics. The model generalizes to 30 unseen sites (∼100 km) across a 5‐year spatiotemporal holdout period (2020–2025), achieving a median Root Mean Squared Error of 12.92 m, producing forecasts without site‐specific calibration. The model has difficulty in regions with complex coastal morphodynamics, such as river entrances and coastal structures that introduce more nuanced relationships that are not explicitly accounted for. Feature analysis further reveals that site‐specific information that contextualizes the primary forcing relationship between waves and shoreline response is essential for generalization. Our findings demonstrate that end‐to‐end deep learning provides a viable pathway for developing regionally trained, generalized shoreline models, and motivates further exploration into the possibility of multi‐regional models.
- 13 February 2026
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 3 (1)
Open AccessA Generalized Data‐Driven Shoreline Model at the Regional Scale
Shoreline models underpin long‐term coastal planning and hazard mitigation but are often limited by site‐specific calibration and simplified process representations. While machine learning (ML) approaches have recently matched or exceeded the skill of traditional models in deterministic shoreline forecasting, they have yet to demonstrate generalization across diverse coastal environments. In this study, we introduce a generalized, data‐driven regional‐scale shoreline model trained across 2,000 km of the southeast Australian coastline. The model adapts the Temporal Fusion Transformer (TFT) architecture, combining a sequence‐to‐sequence Long Short‐Term Memory (LSTM) module with masked multi‐head self‐attention to learn both short‐ and long‐range temporal dependencies. Static covariates unique to each transect are integrated via gating mechanisms that condition the learned relationship between wave action and shoreline response, enabling outputs to adapt to local site characteristics. The model generalizes to 30 unseen sites (∼100 km) across a 5‐year spatiotemporal holdout period (2020–2025), achieving a median Root Mean Squared Error of 12.92 m, producing forecasts without site‐specific calibration. The model has difficulty in regions with complex coastal morphodynamics, such as river entrances and coastal structures that introduce more nuanced relationships that are not explicitly accounted for. Feature analysis further reveals that site‐specific information that contextualizes the primary forcing relationship between waves and shoreline response is essential for generalization. Our findings demonstrate that end‐to‐end deep learning provides a viable pathway for developing regionally trained, generalized shoreline models, and motivates further exploration into the possibility of multi‐regional models.
- Open Access
- 12 September 2025
- Published by Wiley
A Generalised Data-Driven Shoreline Model at the Regional Scale
Shoreline models underpin long-term coastal planning and hazard mitigation but are often limited by site-specific calibration and simplified process representations. While machine learning (ML) approaches have recently matched or exceeded the skill of traditional models in deterministic shoreline forecasting, they have yet to demonstrate generalisation across diverse coastal environments. In this study, we introduce a generalised, data-driven regional-scale shoreline model trained across 2,000 km of the southeast Australian coastline. The model adapts the Temporal Fusion Transformer (TFT) architecture, combining a sequence-to-sequence Long Short Term Memory (LSTM) module with masked multi-head self-attention to learn both short- and long-range temporal dependencies. Static covariates unique to each transect are integrated via gating mechanisms that condition the learned relationship between wave action and shoreline response, enabling outputs to adapt to local site characteristics. The model generalises to 30 unseen sites (~100 km) across a five-year spatiotemporal holdout period (2020-2025), achieving a median RMSE of 12.92 m, producing forecasts without site-specific calibration. The model has difficulty in regions with complex coastal morphodynamics, like river entrances and coastal structures that introduce more nuanced relationships that are not explicitly accounted for. Feature analysis further reveals that site-specific information that contextualises the primary forcing relationship between waves and shoreline response is essential for generalisation. Our findings demonstrate that end-to-end deep learning offers a viable path toward capturing large-scale shoreline dynamics and provides early insight into what may be required to realise unified modelling frameworks capable of operating at regional or even global scales.
- 12 September 2025
- Published by Wiley
Open AccessA Generalised Data-Driven Shoreline Model at the Regional Scale
Shoreline models underpin long-term coastal planning and hazard mitigation but are often limited by site-specific calibration and simplified process representations. While machine learning (ML) approaches have recently matched or exceeded the skill of traditional models in deterministic shoreline forecasting, they have yet to demonstrate generalisation across diverse coastal environments. In this study, we introduce a generalised, data-driven regional-scale shoreline model trained across 2,000 km of the southeast Australian coastline. The model adapts the Temporal Fusion Transformer (TFT) architecture, combining a sequence-to-sequence Long Short Term Memory (LSTM) module with masked multi-head self-attention to learn both short- and long-range temporal dependencies. Static covariates unique to each transect are integrated via gating mechanisms that condition the learned relationship between wave action and shoreline response, enabling outputs to adapt to local site characteristics. The model generalises to 30 unseen sites (~100 km) across a five-year spatiotemporal holdout period (2020-2025), achieving a median RMSE of 12.92 m, producing forecasts without site-specific calibration. The model has difficulty in regions with complex coastal morphodynamics, like river entrances and coastal structures that introduce more nuanced relationships that are not explicitly accounted for. Feature analysis further reveals that site-specific information that contextualises the primary forcing relationship between waves and shoreline response is essential for generalisation. Our findings demonstrate that end-to-end deep learning offers a viable path toward capturing large-scale shoreline dynamics and provides early insight into what may be required to realise unified modelling frameworks capable of operating at regional or even global scales.
- Open Access
- 23 June 2025
- Published by Elsevier in Water Research
- Vol. 285, 124095
A water sensitive urban design (WSUD) planning framework for catchment-scale urban pluvial flood mitigation targets
- 23 June 2025
- Published by Elsevier in Water Research
- Vol. 285, 124095
Open AccessA water sensitive urban design (WSUD) planning framework for catchment-scale urban pluvial flood mitigation targets
- Open Access
- 29 May 2025
- Published by Coastal Engineering Research Council in Coastal Engineering Proceedings
- No. 38,p. 44
A MIXTURE OF EXPERTS APPROACH COMBINING PHYSICS INFORMED AND MACHINE LEARNING SHORELINE MODELS
Understanding the drivers, as well as the ability to predict sandy shoreline change, is of primary interest to coastal engineers and managers, alike. As machine learning has increasingly gained prominence in the field of environmental science, there has also been a growing emphasis on the need for models that are more interpretable. Striking the right balance between model performance and interpretability has become increasingly nuanced. After all, establishing trust in a model necessitates a degree of comprehension regarding its underlying decision-making process. One path forward lies in the co-evolution of both methods to arrive at a modelling approach that retains interpretability and also captures the relevant dynamics to make accurate forecasts. In this study we develop several models combining multiple approaches designed to promote interpretability while retaining high predictive performance. Here we hypothesize that a smart mixture gate informed by prior knowledge will allow more flexibility in capturing the different modes of shoreline change which make up the overall response.
- 29 May 2025
- Published by Coastal Engineering Research Council in Coastal Engineering Proceedings
- No. 38,p. 44
Open AccessA MIXTURE OF EXPERTS APPROACH COMBINING PHYSICS INFORMED AND MACHINE LEARNING SHORELINE MODELS
Understanding the drivers, as well as the ability to predict sandy shoreline change, is of primary interest to coastal engineers and managers, alike. As machine learning has increasingly gained prominence in the field of environmental science, there has also been a growing emphasis on the need for models that are more interpretable. Striking the right balance between model performance and interpretability has become increasingly nuanced. After all, establishing trust in a model necessitates a degree of comprehension regarding its underlying decision-making process. One path forward lies in the co-evolution of both methods to arrive at a modelling approach that retains interpretability and also captures the relevant dynamics to make accurate forecasts. In this study we develop several models combining multiple approaches designed to promote interpretability while retaining high predictive performance. Here we hypothesize that a smart mixture gate informed by prior knowledge will allow more flexibility in capturing the different modes of shoreline change which make up the overall response.
- Open Access
- 18 March 2025
- Published by Copernicus GmbH
Streamflow Response to Tension Water Storage Capacity Distributions in a Large Sample Perspective
Accurately characterizing the spatial variability of tension water storage capacity (TWC) within a catchment is challenging due to limited in-situ hydrologic data availability. Conventional conceptual rainfall-runoff models typically rely on an empirically specified TWC distribution. However, this empirical distribution lacks a physical foundation and fails to effectively redistribute critical hydrologic components, such as local capacity and contributing area, to real-world contexts. To overcome this limitation, the topographic wetness index (TWI) and its generalized form (GTWI) are introduced to bridge local topographic information with hydrologic components. Four TWC distribution curves are contrived based on the empirical parabolic distribution, empirical linear distribution, TWI, and GTWI, respectively. The effects of these alternate TWC distributions on streamflow are investigated within the framework of the HYdrologic MODel (HYMOD) across 460 Australian catchments. The results illustrate that the GTWI-based HYMOD (GTHYMOD) outperforms other models in terms of daily streamflow, with high Kling-Gupta Efficiency (KGE) attained in 74.8% of the study catchments during the validation period. The eastern coast of Australian catchments presents a superior streamflow performance compared to that in the western coast. GTHYMOD demonstrates its superiority in characterizing spatial variability, an aspect HYMOD lacked. This study has the potential to refine the empirical TWC distribution from a physical perspective and advance our comprehension of underlying hydrologic behaviors.
- 18 March 2025
- Published by Copernicus GmbH
Open AccessStreamflow Response to Tension Water Storage Capacity Distributions in a Large Sample Perspective
Accurately characterizing the spatial variability of tension water storage capacity (TWC) within a catchment is challenging due to limited in-situ hydrologic data availability. Conventional conceptual rainfall-runoff models typically rely on an empirically specified TWC distribution. However, this empirical distribution lacks a physical foundation and fails to effectively redistribute critical hydrologic components, such as local capacity and contributing area, to real-world contexts. To overcome this limitation, the topographic wetness index (TWI) and its generalized form (GTWI) are introduced to bridge local topographic information with hydrologic components. Four TWC distribution curves are contrived based on the empirical parabolic distribution, empirical linear distribution, TWI, and GTWI, respectively. The effects of these alternate TWC distributions on streamflow are investigated within the framework of the HYdrologic MODel (HYMOD) across 460 Australian catchments. The results illustrate that the GTWI-based HYMOD (GTHYMOD) outperforms other models in terms of daily streamflow, with high Kling-Gupta Efficiency (KGE) attained in 74.8% of the study catchments during the validation period. The eastern coast of Australian catchments presents a superior streamflow performance compared to that in the western coast. GTHYMOD demonstrates its superiority in characterizing spatial variability, an aspect HYMOD lacked. This study has the potential to refine the empirical TWC distribution from a physical perspective and advance our comprehension of underlying hydrologic behaviors.
- Open Access
- 20 January 2025
- Published by Copernicus GmbH
Identifying input and output data errors in the calibration of a water quality model using Bayesian error analysis with reordering (BEAR) method
The presence of errors in water quality and hydrologic variables can significantly impair the calibration of water quality models. To enhance the estimation of model parameters, it is important to accurately identify data errors during the calibration process. However, this task is challenging due to the complex interactions between model parameter uncertainty and data uncertainty. Existing methods for incorporating data uncertainty in model calibration have limitations, such as high-dimensional computation or the inability to handle stochastic errors.To address these challenges, a novel method called Bayesian Error Analysis with Reordering (BEAR) has been developed. Given that the data uncertainty arises from the data itself and is independent of the model calibration or simulation, the cumulative distribution function (CDF) of the data error can be estimated ahead and regarded as the prior information of Bayesian inference. Then the values of data error series only depend on their ranks in the CDF. BEAR method transforms the values of data error series into their ranks in the CDF. This transformation enables the effective identification of input and/or output data errors in water quality calibration.The innovation of the BEAR method can be attributed to several key aspects:1) Modification of the secant method to handle the non-linear transformation from input to output, ensuring the correspondence between the rank of input data error and the residual error of the model.2) Decomposition of model simulations to calculate the delay between each input and its corresponding output.3) Utilization of the Autoregressive model to account for the correlation of residual errors.4) Selection of an appropriate updating logic to minimize the compensation effects among multiple sources of data uncertainty.Overall, the BEAR method demonstrates flexibility and adaptability to various environmental modelling scenarios, making it a valuable tool for improving model specification under conditions of data uncertainty.
- 20 January 2025
- Published by Copernicus GmbH
Open AccessIdentifying input and output data errors in the calibration of a water quality model using Bayesian error analysis with reordering (BEAR) method
The presence of errors in water quality and hydrologic variables can significantly impair the calibration of water quality models. To enhance the estimation of model parameters, it is important to accurately identify data errors during the calibration process. However, this task is challenging due to the complex interactions between model parameter uncertainty and data uncertainty. Existing methods for incorporating data uncertainty in model calibration have limitations, such as high-dimensional computation or the inability to handle stochastic errors.To address these challenges, a novel method called Bayesian Error Analysis with Reordering (BEAR) has been developed. Given that the data uncertainty arises from the data itself and is independent of the model calibration or simulation, the cumulative distribution function (CDF) of the data error can be estimated ahead and regarded as the prior information of Bayesian inference. Then the values of data error series only depend on their ranks in the CDF. BEAR method transforms the values of data error series into their ranks in the CDF. This transformation enables the effective identification of input and/or output data errors in water quality calibration.The innovation of the BEAR method can be attributed to several key aspects:1) Modification of the secant method to handle the non-linear transformation from input to output, ensuring the correspondence between the rank of input data error and the residual error of the model.2) Decomposition of model simulations to calculate the delay between each input and its corresponding output.3) Utilization of the Autoregressive model to account for the correlation of residual errors.4) Selection of an appropriate updating logic to minimize the compensation effects among multiple sources of data uncertainty.Overall, the BEAR method demonstrates flexibility and adaptability to various environmental modelling scenarios, making it a valuable tool for improving model specification under conditions of data uncertainty.