Uncovering Flooding Mechanisms Across the Contiguous United States Through Interpretive Deep Learning on Representative Catchments

Water Resources Research, Journal Year: 2021, Volume and Issue: 58(1)

Published: Dec. 27, 2021

Abstract Long short‐term memory (LSTM) networks represent one of the most prevalent deep learning (DL) architectures in current hydrological modeling, but they remain black boxes from which process understanding can hardly be obtained. This study aims to demonstrate potential interpretive DL gaining scientific insights using flood prediction across contiguous United States (CONUS) as a case study. Two interpretation methods were adopted decipher machine‐captured patterns and inner workings LSTM networks. The by expected gradients method revealed three distinct input‐output relationships learned LSTM‐based runoff models 160 individual catchments. These correspond flood‐inducing mechanisms—snowmelt, recent rainfall, historical rainfall—that account for 10.1%, 60.9%, 29.0% 20,908 flow peaks identified data set, respectively. Single flooding mechanisms dominate 70.7% investigated catchments (11.9% snowmelt‐dominated, 34.4% rainfall‐dominated, 24.4% rainfall‐dominated mechanisms), remaining 29.3% have mixed mechanisms. spatial variability dominant reflects catchments' geographic climatic conditions. Moreover, additive decomposition unveils how network behaves differently retaining discarding information when emulating different types floods. Information inputs within previous time steps partially stored predict snowmelt‐induced rainfall‐induced floods, while only is retained. Overall, this provides new perspective processes extremes demonstrates prospect artificial intelligence‐assisted discovery future.

Language: Английский

---

Reconciling disagreement on global river flood changes in a warming climate

Nature Climate Change, Journal Year: 2022, Volume and Issue: 12(12), P. 1160 - 1167

Published: Nov. 28, 2022

Language: Английский

---

Understanding the relationship between rainfall and flood probabilities through combined intensity-duration-frequency analysis

Journal of Hydrology, Journal Year: 2021, Volume and Issue: 602, P. 126759 - 126759

Published: Aug. 2, 2021

The aim of this paper is to explore how rainfall mechanisms and catchment characteristics shape the relationship between flood probabilities. We propose a new approach comparing intensity-duration-frequency statistics maximum annual with those streamflow in order infer behavior for runoff extremes. calibrate parsimonious scaling models data from 314 rain gauges 428 stream Austria, analyze spatial patterns resulting distributions model parameters. Results indicate that extremes tend be more variable dry lowland catchments dominated by convective than mountainous where higher are mainly orographic. Flood frequency curves always steeper corresponding exception glaciated catchments. Based on proposed combined we elasticities as percent change discharge 1% extreme through quantiles. In wet catchments, unity, i.e. have similar steepness, due persistently high soil moisture levels. much higher, implying floods rainfall, which interpreted terms skewed event coefficients. While regional differences can attributed both dominating characteristics, our results suggest controls. With increasing return period, towards consistent various generation concepts. Our findings may useful process-based extrapolation climate impact studies, further studies encouraged tail elasticities.

Language: Английский

---

Generalization of an Encoder-Decoder LSTM model for flood prediction in ungauged catchments

Journal of Hydrology, Journal Year: 2022, Volume and Issue: 614, P. 128577 - 128577

Published: Oct. 30, 2022

Flood prediction in ungauged catchments is usually conducted by hydrological models that are parameterized based on nearby and similar gauged catchments. As an alternative to this process-based modelling, deep learning (DL) have demonstrated their ability for (PUB) with high efficiency. Catchment characteristics, the number of catchments, level hydroclimatic heterogeneity training dataset used model regionalization can directly affect model’s performance. Here, we study generalization a DL these factors applying Encoder-Decoder Long Short-Term Memory neural network 6-hour lead-time runoff 35 mountainous China. By varying available settings different datasets, namely local, regional, PUB models, evaluated our model. We found both quantity (i.e. available) important improving performance context, due data synergy effect. The assessment sensitivity catchment characteristics showed mainly correlated local hydro-climatic conditions; more arid region, likely it poor results suggest regional ED-LSTM promising method predict streamflow from rainfall inputs PUB, outline need preparing representative dataset.

Language: Английский

---

Optimal Postprocessing Strategies With LSTM for Global Streamflow Prediction in Ungauged Basins

Water Resources Research, Journal Year: 2023, Volume and Issue: 59(7)

Published: July 1, 2023

Abstract Streamflow prediction in ungauged basins (PUB) is challenging, and Long Short‐Term Memory (LSTM) widely used to for such predictions, owing its excellent migration performance. Traditional LSTM forced by meteorological data catchment attribute barely highlight the optimum integration strategy from data‐rich ones. In this study, we experimented with 1,897 global catchments found that LSTM‐corrected Global Hydrological Models (GHMs) outperformed uncorrected GHMs, improving median Nash‐Sutcliff efficiency (NSE) 0.03 0.66. Notably, there was a large gap between traditional modeling autoregressive basins, GHM‐forced were an effective way close basins. The spatial heterogeneity of performance mainly influenced three metrics (dryness, leaf area index latitude), which described hydrological similarity among catchments. Weaker continental results larger variability LSTM, best Siberia (NSE, 0.54) worst North America 0.10). However, significantly improved 0.63) when considered. This study stressed advantages due significance should be attached similarities improve

Language: Английский

---

Compounding effects in flood drivers challenge estimates of extreme river floods

Science Advances, Journal Year: 2024, Volume and Issue: 10(13)

Published: March 27, 2024

Estimating river flood risks under climate change is challenging, largely due to the interacting and combined influences of various flood-generating drivers. However, a more detailed quantitative analysis such compounding effects implications their interplay remains underexplored on large scale. Here, we use explainable machine learning disentangle between drivers quantify importance for different magnitudes across thousands catchments worldwide. Our findings demonstrate ubiquity in many floods. Their often increases with magnitude, but strength this increase varies basis catchment conditions. Traditional might underestimate extreme hazards where contribution strongly magnitude. Overall, our study highlights need carefully incorporate risk assessment improve estimates

Language: Английский

---

Catchment characterization: Current descriptors, knowledge gaps and future opportunities

Earth-Science Reviews, Journal Year: 2024, Volume and Issue: 252, P. 104739 - 104739

Published: March 8, 2024

The ability to characterize hydrologically relevant differences between places is at the core of our science. A common way quantitatively hydrological catchments through use descriptors that summarize physical aspects system, typically by aggregating heterogeneous geospatial information into a single number. Such capture various facets catchment functioning and structure, identify similarity or dissimilarity among catchments, transfer unobserved locations. However, so far there no agreement on how should be selected, aggregated, evaluated. Even worse, little known about existence potential biases in current practices catchments. In this systematic review, we analyze 742 research articles published 1967 2021 provide categorized overview historical characterization (i.e., data sources, aggregation evaluation methods) science related disciplines. We uncover substantial characterization: (1) only 16% analyzed studies are dry environments, even though such environments cover 42% global land surface, suggesting most tailored represent energy-limited potentially less effective water-limited environments; (2) 30% subsurface features for despite dominance flow; (3) 4% 9% aggregated spatially- vertically-differentiated way, respectively, while majority simple averages do not account hydrologically-relevant variabilities within catchments; (4) 25% all evaluate usefulness descriptors, none quantifies their uncertainty. demonstrate effects these effectively functional behavior with illustrative examples. Finally, suggest possible ways derive more robust, comprehensive meaningful descriptors.

Language: Английский

---

Streamflow prediction in ungauged catchments through use of catchment classification and deep learning

Journal of Hydrology, Journal Year: 2024, Volume and Issue: 639, P. 131638 - 131638

Published: July 3, 2024

Language: Английский

---

Future Change in Urban Flooding Using New Convection‐Permitting Climate Projections

Water Resources Research, Journal Year: 2024, Volume and Issue: 60(1)

Published: Jan. 1, 2024

Abstract Rainfall intensity in the United Kingdom is projected to increase under climate change with significant implications for rainfall‐driven (combined pluvial and fluvial) flooding. In UK, current recommended best practice estimating changes flood hazard involves applying a simple percentage uplift spatially uniform catchment rainfall, despite known importance of spatial temporal characteristics rainfall generation floods. The UKCP Local Convective Permitting Model (CPM) has first time provided capacity assess using hourly, 2.2 km CPM precipitation data that varies space time. Here, we use an event set ∼13,500 events across three epochs (1981–2000, 2021–2040, 2061–2080) simulate flooding LISFLOOD‐FP hydrodynamic model at 20 m resolution over 750 2 area Bristol Bath, UK. We find both approaches indicate near‐term (2021–2040) future (2061–2080) change. However, produces markedly higher estimates when compared approach, ranging from 19% 49% depending on return period. This suggests including full spatiotemporal variability its modeling critical risk assessment.

Language: Английский

---

Analysis of pollutant dispersion patterns in rivers under different rainfall based on an integrated water-land model

Journal of Environmental Management, Journal Year: 2024, Volume and Issue: 354, P. 120314 - 120314

Published: Feb. 23, 2024

Language: Английский

---