Search for: All records

Abstract Artificial neural networks are increasingly used for geophysical modeling to extract complex nonlinear patterns from geospatial data. However, it is difficult to understand how networks make predictions, limiting trust in the model, debugging capacity, and physical insights. EXplainable Artificial Intelligence (XAI) techniques expose how models make predictions, but XAI results may be influenced by correlated features. Geospatial data typically exhibit substantial autocorrelation. With correlated input features, learning methods can produce many networks that achieve very similar performance (e.g., arising from different initializations). Since the networks capture different relationships, their attributions can vary. Correlated features may also cause inaccurate attributions because XAI methods typically evaluate isolated features, whereas networks learn multifeature patterns. Few studies have quantitatively analyzed the influence of correlated features on XAI attributions. We use a benchmark framework of synthetic data with increasingly strong correlation, for which the ground truth attribution is known. For each dataset, we train multiple networks and compare XAI-derived attributions to the ground truth. We show that correlation may dramatically increase the variance of the derived attributions, and investigate the cause of the high variance: is it because different trained networks learn highly different functions or because XAI methods become less faithful in the presence of correlation? Finally, we show XAI applied to superpixels, instead of single grid cells, substantially decreases attribution variance. Our study is the first to quantify the effects of strong correlation on XAI, to investigate the reasons that underlie these effects, and to offer a promising way to address them. 

Accurate and timely water level predictions are essential for effective shoreline and coastal ecosystem management. As sea levels rise, the frequency and severity of coastal inundation events are increasing, causing significant societal and economic impacts. Predicting these events with sufficient lead time is essential for decision-makers to mitigate economic losses and protect coastal communities. While machine learning methods have been developed to predict water levels at specific sites, there remains a need for more generalized models that perform well across diverse locations. This study presents a robust deep learning model for predicting water levels at multiple tide gauge locations along the Gulf of Mexico, including the open coast, embayments, and ship channels, all near major ports. The selected architecture, Seq2Seq, achieves significant improvements over the existing literature. It meets the National Oceanic and Atmospheric Administration’s (NOAA) operational criterion, with the percentage of predictions within 15 cm for lead times up to 108 h at the tide gauges of Port Isabel (92.2%) and Rockport (90.4%). These results represent a significant advancement over current models typically failing to meet NOAA’s standard beyond 48 h. This highlights the potential of deep learning models to improve water level predictions, offering crucial support for coastal management and flood mitigation. 

By improving the prediction, understanding, and communication of powerful events in the atmosphere and ocean, artificial intelligence can revolutionize how communities respond to climate change. 

Mechanisms that generate subseasonal (1-2 months) events of sea level rise along the western Gulf Coast are investigated using the data collected by a dense tide gauge network: Texas Coastal Ocean Observation Network (TCOON) and National Water Level Observation Network (NWLON), satellite altimetry, and high-resolution (0.08°) ocean reanalysis product. In particular, the role of Loop Current and eddy shedding in generating the extreme sea level rise along the coast is emphasized. The time series of sea level anomalies along the western portion of the Gulf Coast derived from the TCOON and NWLON tide gauge data indicate that a subseasonal sea level rise which exceeds 15 cm is observed once in every 2-5 years. Based on the analysis of satellite altimetry data and high-resolution ocean reanalysis product, it is found that most of such extreme subseasonal events are originated from the anti-cyclonic (warm-core) eddy separated from the Loop Current which propagates westward. A prominent sea level rise is generated when the eddy reaches the western Gulf Coast, which occurs about 6-8 months after the formation of strong anti-cyclonic eddy in the central Gulf of Mexico. The results demonstrate that the accurate prediction of subseasonal sea level rise events along the Gulf Coast with the lead time of several months require a full description of large-scale ocean dynamical processes in the entire Gulf of Mexico including the characteristics of eddies separated from the Loop Current.