An official website of the United States government
Abstract Recent advances in explainable artificial intelligence (XAI) methods show promise for understanding predictions made by machine learning (ML) models. XAI explains how the input features are relevant or important for the model predictions. We train linear regression (LR) and convolutional neural network (CNN) models to make 1-day predictions of sea ice velocity in the Arctic from inputs of present-day wind velocity and previous-day ice velocity and concentration. We apply XAI methods to the CNN and compare explanations to variance explained by LR. We confirm the feasibility of using a novel XAI method [i.e., global layerwise relevance propagation (LRP)] to understand ML model predictions of sea ice motion by comparing it to established techniques. We investigate a suite of linear, perturbation-based, and propagation-based XAI methods in both local and global forms. Outputs from different explainability methods are generally consistent in showing that wind speed is the input feature with the highest contribution to ML predictions of ice motion, and we discuss inconsistencies in the spatial variability of the explanations. Additionally, we show that the CNN relies on both linear and nonlinear relationships between the inputs and uses nonlocal information to make predictions. LRP shows that wind speed over land is highly relevant for predicting ice motion offshore. This provides a framework to show how knowledge of environmental variables (i.e., wind) on land could be useful for predicting other properties (i.e., sea ice velocity) elsewhere. Significance StatementExplainable artificial intelligence (XAI) is useful for understanding predictions made by machine learning models. Our research establishes trustability in a novel implementation of an explainable AI method known as layerwise relevance propagation for Earth science applications. To do this, we provide a comparative evaluation of a suite of explainable AI methods applied to machine learning models that make 1-day predictions of Arctic sea ice velocity. We use explainable AI outputs to understand how the input features are used by the machine learning to predict ice motion. Additionally, we show that a convolutional neural network uses nonlinear and nonlocal information in making its predictions. We take advantage of the nonlocality to investigate the extent to which knowledge of wind on land is useful for predicting sea ice velocity elsewhere.
more »
« less
Machine Learning for Daily Forecasts of Arctic Sea Ice Motion: An Attribution Assessment of Model Predictive Skill
Physics-based simulations of Arctic sea ice are highly complex, involving transport between different phases, length scales, and time scales. Resultantly, numerical simulations of sea ice dynamics have a high computational cost and model uncertainty. We employ data-driven machine learning (ML) to make predictions of sea ice motion. The ML models are built to predict present-day sea ice velocity given present-day wind velocity and previous-day sea ice concentration and velocity. Models are trained using reanalysis winds and satellite-derived sea ice properties. We compare the predictions of three different models: persistence (PS), linear regression (LR), and a convolutional neural network (CNN). We quantify the spatiotemporal variability of the correlation between observations and the statistical model predictions. Additionally, we analyze model performance in comparison to variability in properties related to ice motion (wind velocity, ice velocity, ice concentration, distance from coast, bathymetric depth) to understand the processes related to decreases in model performance. Results indicate that a CNN makes skillful predictions of daily sea ice velocity with a correlation up to 0.81 between predicted and observed sea ice velocity, while the LR and PS implementations exhibit correlations of 0.78 and 0.69, respectively. The correlation varies spatially and seasonally: lower values occur in shallow coastal regions and during times of minimum sea ice extent. LR parameter analysis indicates that wind velocity plays the largest role in predicting sea ice velocity on 1-day time scales, particularly in the central Arctic. Regions where wind velocity has the largest LR parameter are regions where the CNN has higher predictive skill than the LR.
more »
« less
- PAR ID:
- 10481308
- Publisher / Repository:
- AMS
- Date Published:
- Journal Name:
- Artificial Intelligence for the Earth Systems
- Volume:
- 2
- Issue:
- 4
- ISSN:
- 2769-7525
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Although standard statistical methods and climate models can simulate and predict sea-ice changes well, it is still very hard to distinguish some direct and robust factors associated with sea-ice changes from its internal variability and other noises. Here, with long-term observations (38 years from 1980 to 2017), we apply the causal effect networks algorithm to explore the direct precursors of September Arctic sea-ice extent by adjusting the maximal lead time from one to eight months. For lead time of more than three months, June downward longwave radiation flux in the Canadian Arctic Archipelago is the only one precursor. However, for lead time of 1–3 months, August sea-ice concentration in Western Arctic represents the strongest positive correlation with September sea-ice extent, while August sea-ice concentration factors in other regions have weaker influences on the marginal seas. Other precursors include August wind anomalies in the lower latitudes accompanied with an Arctic high pressure anomaly, which induces the sea-ice loss along the Eurasian coast. These robust precursors can be used to improve the seasonal predictions of Arctic sea ice and evaluate the climate models.more » « less
-
Abstract The variability of Arctic sea ice extent (SIE) on interannual and multidecadal time scales is examined in 29 models with historical forcing participating in phase 6 of the Coupled Model Intercomparison Project (CMIP6) and in twentieth-century sea ice reconstructions. Results show that during the historical period with low external forcing (1850–1919), CMIP6 models display relatively good agreement in their representation of interannual sea ice variability (IVSIE) but exhibit pronounced intermodel spread in multidecadal sea ice variability (MVSIE), which is overestimated with respect to sea ice reconstructions and is dominated by model uncertainty in sea ice simulation in the subpolar North Atlantic. We find that this is associated with differences in models’ sensitivity to Northern Hemispheric sea surface temperatures (SSTs). Additionally, we show that while CMIP6 models are generally capable of simulating multidecadal changes in Arctic sea ice from the mid-twentieth century to present day, they tend to underestimate the observed sea ice decline during the early twentieth-century warming (ETCW; 1915–45). These results suggest the need for an improved characterization of the sea ice response to multidecadal climate variability in order to address the sources of model bias and reduce the uncertainty in future projections arising from intermodel spread. Significance StatementThe credibility of Arctic sea ice predictions depends on whether climate models are capable of reproducing changes in the past climate, including patterns of sea ice variability which can mask or amplify the response to global warming. This study aims to better understand how latest-generation global climate models simulate interannual and multidecadal variability of Arctic sea ice relative to available observations. We find that models differ in their representation of multidecadal sea ice variability, which is overall larger than in observations. Additionally, models underestimate the sea ice decline during the period of observed warming between 1915 and 1945. Our results suggest that, to achieve better predictions of Arctic sea ice, the realism of low-frequency sea ice variability in models should be improved.more » « less
-
Abstract Previous findings show that large-scale atmospheric circulation plays an important role in driving Arctic sea ice variability from synoptic to seasonal time scales. While some circulation patterns responsible for Barents–Kara sea ice changes have been identified in previous works, the most important patterns and the role of their persistence remain unclear. Our study uses self-organizing maps to identify nine high-latitude circulation patterns responsible for day-to-day Barents–Kara sea ice changes. Circulation patterns with a high pressure center over the Urals (Scandinavia) and a low pressure center over Iceland (Greenland) are found to be the most important for Barents–Kara sea ice loss. Their opposite-phase counterparts are found to be the most important for sea ice growth. The persistence of these circulation patterns helps explain sea ice variability from synoptic to seasonal time scales. We further use sea ice models forced by observed atmospheric fields (including the surface circulation and temperature) to reproduce observed sea ice variability and diagnose the role of atmosphere-driven thermodynamic and dynamic processes. Results show that thermodynamic and dynamic processes similarly contribute to Barents–Kara sea ice concentration changes on synoptic time scales via circulation. On seasonal time scales, thermodynamic processes seem to play a stronger role than dynamic processes. Overall, our study highlights the importance of large-scale atmospheric circulation, its persistence, and varying physical processes in shaping sea ice variability across multiple time scales, which has implications for seasonal sea ice prediction. Significance StatementUnderstanding what processes lead to Arctic sea ice changes is important due to their significant impacts on the ecosystem, weather, and shipping, and hence our society. A well-known process that causes sea ice changes is atmospheric circulation variability. We further pin down what circulation patterns and underlying mechanisms matter. We identify multiple circulation patterns responsible for sea ice loss and growth to different extents. We find that the circulation can cause sea ice loss by mechanically pushing sea ice northward and bringing warm and moist air to melt sea ice. The two processes are similarly important. Our study advances understanding of the Arctic sea ice variability with important implications for Arctic sea ice prediction.more » « less
-
Recent climate change in the Arctic has been rapid and dramatic, leading to numerous physical and societal consequences. Many studies have investigated these ongoing and projected future changes across a range of climatic variables, but surprisingly little attention has been paid to wind speed, despite its known importance for sea ice motion, ocean wave heights, and coastal erosion. Here we analyzed future trends in Arctic surface wind speed and its relationship with sea ice cover among CMIP5 global climate models. There is a strong anticorrelation between climatological sea ice concentration and wind speed in the early 21st-century reference climate, and the vast majority of models simulate widespread future strengthening of surface winds over the Arctic Ocean (annual multi-model mean trend of up to 0.8 m s−1 or 13%). Nearly all models produce an inverse relationship between projected changes in sea ice cover and wind speed, such that grid cells with virtually total ice loss almost always experience stronger winds. Consistent with the largest regional ice losses during autumn and winter, the greatest increases in future wind speeds are expected during these two seasons, with localized strengthening up to 23%. As in other studies, stronger surface winds cannot be attributed to tighter pressure gradients but rather to some combination of weakened atmospheric stability and reduced surface roughness as the surface warms and melts. The intermodel spread of wind speed changes, as expressed by the two most contrasting model results, appears to stem from differences in the treatment of surface roughness.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1175/AIES-D-23-0004.1
