Denoising Diffusion Probabilistic Models for Coastal Inundation Forecasting

Kazi Ashik Islam, Zakaria Mehrab, Mahantesh Halappanavar, Henning Mortveit, Sridhar Katragadda, Jon Derek Loftis, Madhav Marathe · 2025

Read Denoising Diffusion Probabilistic Models for Coastal Inundation Forecasting by Kazi Ashik Islam, Zakaria Mehrab, Mahantesh Halappanavar, Henning Mortveit, Sridhar Katragadda, Jon Derek Loftis, Madhav Marathe (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-05381.txt.

Introduction

Kazi Ashik Islam and co-authors present a 2025 system for forecasting coastal flooding at fine detail from sparse sensors. Their hosted title frames the work as high resolution prediction from few observations. Flooding threatens coastal towns more each year, yet many places can afford gauges at only a few sites.

The system, named DIFF-SPARSE, is a masked diffusion model that gives odds, not just one forecast. It reads the recent flood history of a site and its neighbors, plus land height data and time cues. A new masking plan in training teaches it to cope with gaps. Tested on Virginia's Eastern Shore at levels up to ninety-five percent missing data, it beat prior methods by up to about sixty-two percent on two forecast scores.

Catalog record

CREATORKazi Ashik Islam, Zakaria Mehrab, Mahantesh Halappanavar, Henning Mortveit, Sridhar Katragadda, Jon Derek Loftis, Madhav Marathe
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-05381

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Denoising Diffusion Probabilistic Models for Coastal Inundation Forecasting Kazi Ashik Islam, Zakaria Mehrab, Mahantesh Halappanavar, Henning Mortveit, Sridhar Katragadda, Jon Derek Loftis, Madhav Marathe Abstract Coastal flooding poses significant risks to communities, necessitating fast and accurate forecasting methods to mitigate potential damage. To approach this problem, we present Diff-Flood , a probabilistic spatiotemporal forecasting method designed based on denoising diffusion models. Diff-Flood predicts inundation level at a location by taking both spatial and temporal context into account. It utilizes inundation levels at neighboring locations and digital elevation data as spatial context. Inundation history from a context time window, together with additional co-variates are used as temporal context. Convolutional neural networks and cross-attention mechanism are then employed to capture the spatiotemporal dynamics in the data. We trained and tested Diff-Flood on coastal inundation data from the Eastern Shore of Virginia, a region highly impacted by coastal flooding. Our results show that, Diff-Flood outperforms existing forecasting methods in terms of prediction performance ( 6 % ​ to ​ 64 % 6\%\text{ to }64\% improvement in terms of two performance metrics) and scalability. 1 Introduction The projection in sea level rise have highlighted the increasing vulnerability of many coastal regions to inundation ( Wuebbles et al. 2017 ) . Among these regions, the East Coast in US is a region of particular policy concern due to its geography and low-lying topography ( Ezer and Atkinson 2014 ) . This region, in recent periods, has been experiencing frequent inundation events ( Ezer 2020 ) , causing disruptions in daily life (e.g. traffic, economic) ( Jacobs et al. 2018 ) . In the future, this situation will only keep declining due to the combined effect of climate change, sea level rise, and urbanization ( Swain et al. 2020 ) . Therefore, forecasting coastal inundation is of utmost importance from a policy standpoint for undertaking risk prevention measurement or safe evacuation of the residents in danger ( Yang et al. 2019 ) . Figure 1: Flooding on Quinby Bridge Road, Accomack County, VA on April 3, 2024 (Data source: Tidewatch Maps ( Tidewatch 2025 ) .) The problem of coastal inundation forecasting came to our attention during conversations with staff at the City of Virginia Beach. The region faces challenges with coastal flooding which causes slow-down and/or road closures; disrupting the day-to-day activities of the inhabitants. A fast and high-resolution flood forecasting model would enable city officials to identify localized inundation risks, resulting in more effective preparedness and response. In this paper, we approach the problem of designing such a model. It is a joint-work of a multidisciplinary team, including computer and data scientists, academic experts with expertise in hydro-dynamic models, members of established coastal resources management centers, and officials at the City of Virginia Beach. The traditional method for predicting inundation involves hydro-dynamical, physics-based models (PBM). Commercially available ( Syme 2001 ; Deltares 2014 ) (e.g., SOBEK, TUFLOW) or open-sourced ( Zhang et al. 2016 ) (e.g, SCHISM) hydro-dynamical models can simulate inundation accurately at a high resolution. However, the computation time associated with these PBMs makes them challenging for real-time forecasting ( Guo and others 2021 ) . Although researchers have investigated Deep-Learning (DL) based time-series prediction methods as surrogates to these PBMs ( Roy et al. 2023 ) , these models are not suitable for accommodating spatial contexts. Existing spatiotemporal forecasting methods, on the other hand, struggle to scale in high-resolution forecasting tasks. Motivated by the above observations, we explore the Diffusion Model ( Sohl-Dickstein et al. 2015 ) in the context of high-resolution coastal inundation forecasting task. The primary reason for investigating diffusion models is its ability to utilize conditioning contexts of various dimensions ( Rombach et al. 2022 ; Zhang and others 2023 ) and its potential to model physical properties ( Ahmad and others 2024 ; Amram and Pedro 2023 ) . Originally developed for image generation, diffusion models have been adapted for time-series forecasting, achieving state-of-the-art performance, as demonstrated by Rasul et al. ( Rasul et al. 2021 ). However, their approach does not incorporate spatial context. More recently, ( Wen et al. 2023 ) proposed a probabilistic spatiotemporal graph forecasting method based on denoising diffusion models. While promising, such methods face scalability challenges when applied to large-scale/high-resolution inundation forecasting tasks. Using the coastal inundation of the Eastern Shore as case study, we address these challenges by making the following contributions in this paper. • We formulate the problem of coastal inundation forecasting under the spatiotemporal forecasting framework and introduce Diff-Flood as a novel method for approaching the problem. It can utilize both spatial and temporal context to generate spatiotemporal forecasts at fine spatial resolution. • Extensive evaluation of our method against several baseline methods on inundation data of the Virginia Eastern Shore underscores the superiority of our proposed method and its aspect of being used as a decision-making tool for policymakers. Specifically, we observed 6\% to 64\% improvement in prediction performance, in terms of two performance metrics (Normalized Average Continuous Ranked Probability Score, and Normalized Root Mean Square Error). An ablation study of our model is performed to highlight the importance of its components. • Furthermore, we demonstrate the scalability of our model compared to other baselines in large settings, underscoring its utility in real-time forecasting in large-scale scenarios. It is worth highlighting that our model can be used to answer inundation scenario-based queries relevant to policymakers. For example, queries of the form: “What is the probability that a particular region will get inundated by 1 feet within the next 12 hours?” , can be answered by using our model through probabilistic forecasting, see section 7 . 2 Related Work We present an abridged version of the most relevant literature in this section. The supplementary material of this paper covers the literature more extensively. Hydrology Models: Traditionally, various Physics Based hydrodynamic Models (PBM) have been employed to simulate inundation. SOBEK ( Deltares 2014 ) is a 2D model that solves a Saint-Venant flow equation and shallow water equation to determine the water level of rectangular grids. TELEMAC-2D ( Vu et al. 2015 ) uses a similar method but is more suitable for coastal areas. Polymorphic models like SCHISM ( Zhang et al. 2016 ) can capture water flow at very high resolution. However, the computation time associated with solving these complex equations at high spatial resolution makes these PBMs infeasible for real-time forecasting. Deep Learning: DL methods are a popular choice in surrogating hydrodynamic models. Regression-based methods ( Zahura and Goodall 2022 ) , recurrent neural networks ( Roy et al. 2023 ) and hybrid ML-based methods ( Zanchetta and Coulibaly 2022 ) are some of the methods that have been attempted as possible surrogates. However, these models cannot inherently take spatial context into account. The DCRNN method proposed by ( Li et al. 2017 ) for traffic forecasting is the state-of-the-art DL method for spatiotemporal point forecasting, where bidirectional random walk is used to capture spatial dependency and Seq2Seq architecture is used to capture temporal dependency. Diffusion Model The field of generative modeling has been dominated by the diffusion model for quite some time. Although it had primarily been studied for image synthesis ( Ho et al.

Read the complete text