Search for: All records

Abstract The rapid intensification (RI) of tropical cyclones (TC), defined here as an intensity increase of ≥ 30 kt in 24 hours, is a difficult but important forecasting problem. Operational RI forecasts have considerably improved since the late 2000s, largely thanks to better statistical models, including machine learning (ML). Most ML applications use scalars from the Statistical Hurricane Intensity Prediction Scheme (SHIPS) development dataset as predictors, describing the TC history, near-TC environment, and satellite presentation of the TC. More recent ML applications use convolutional neural networks (CNN), which can ingest full satellite images (or time series of images) and freely “decide” which spatiotemporal features are important for RI. However, two questions remain unanswered: (1) Does image convolution significantly improve RI skill? (2) What strategies do CNNs use for RI prediction – and can we gain new insights from these strategies? We use an ablation experiment to answer the first question and explainable artificial intelligence (XAI) to answer the second. Convolution leads to only a small performance gain, likely because, as revealed by XAI, the CNN’s main strategy uses image features already well described in scalar predictors used by pre-existing RI models. This work makes three additional contributions to the literature: (1) NNs with SHIPS data outperform pre-existing models in some aspects; (2) NNs provide well calibrated uncertainty quantification (UQ), while pre-existing models have no UQ; (3) the NN without SHIPS data performs surprisingly well and is fairly independent of pre-existing models, suggesting its potential value in an operational ensemble. 

Abstract Pure artificial intelligence (AI)-based weather prediction (AIWP) models have made waves within the scientific community and the media, claiming superior performance to numerical weather prediction (NWP) models. However, these models often lack impactful output variables such as precipitation. One exception is Google DeepMind’s GraphCast model, which became the first mainstream AIWP model to predict precipitation, but performed only limited verification. We present an analysis of the ECMWF’s Integrated Forecasting System (IFS)-initialized (GRAP_IFS) and the NCEP’s Global Forecast System (GFS)-initialized (GRAP_GFS) GraphCast precipitation forecasts over the contiguous United States and compare to results from the GFS and IFS models using 1) grid-based, 2) neighborhood, and 3) object-oriented metrics verified against the fifth major global reanalysis produced by ECMWF (ERA5) and the NCEP/Environmental Modeling Center (EMC) stage IV precipitation analysis datasets. We affirmed that GRAP_GFSand GRAP_IFSperform better than the GFS and IFS in terms of root-mean-square error and stable equitable errors in probability space, but the GFS and IFS precipitation distributions more closely align with the ERA5 and stage IV distributions. Equitable threat score also generally favored GraphCast, particularly for lower accumulation thresholds. Fractions skill score for increasing neighborhood sizes shows greater gains for the GFS and IFS than GraphCast, suggesting the NWP models may have a better handle on intensity but struggle with the location. Object-based verification for GraphCast found positive area biases at low accumulation thresholds and large negative biases at high accumulation thresholds. GRAP_GFSsaw similar performance gains to GRAP_IFSwhen compared to their NWP counterparts, but initializing with the less familiar GFS conditions appeared to lead to an increase in light precipitation. Significance StatementPure artificial intelligence (AI)-based weather prediction (AIWP) has exploded in popularity with promises of better performance and faster run times than numerical weather prediction (NWP) models. However, less attention has been paid to their capability to predict impactful, sensible weather like precipitation, precipitation type, or specific meteorological features. We seek to address this gap by comparing precipitation forecast performance by an AI model called GraphCast to the Global Forecast System (GFS) and the Integrated Forecasting System (IFS) NWP models. While GraphCast does perform better on many verification metrics, it has some limitations for intense precipitation forecasts. In particular, it less frequently predicts intense precipitation events than the GFS or IFS. Overall, this article emphasizes the promise of AIWP while at the same time stresses the need for robust verification by domain experts. 

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. 

Abstract AI-based algorithms are emerging in many meteorological applications that produce imagery as output, including for global weather forecasting models. However, the imagery produced by AI algorithms, especially by convolutional neural networks (CNNs), is often described as too blurry to look realistic, partly because CNNs tend to represent uncertainty as blurriness. This blurriness can be undesirable since it might obscure important meteorological features. More complex AI models, such as Generative AI models, produce images that appear to be sharper. However, improved sharpness may come at the expense of a decline in other performance criteria, such as standard forecast verification metrics. To navigate any trade-off between sharpness and other performance metrics it is important to quantitatively assess those other metrics along with sharpness. While there is a rich set of forecast verification metrics available for meteorological images, none of them focus on sharpness. This paper seeks to fill this gap by 1) exploring a variety of sharpness metrics from other fields, 2) evaluating properties of these metrics, 3) proposing the new concept of Gaussian Blur Equivalence as a tool for their uniform interpretation, and 4) demonstrating their use for sample meteorological applications, including a CNN that emulates radar imagery from satellite imagery (GREMLIN) and an AI-based global weather forecasting model (GraphCast). 

Abstract Numerous artificial intelligence-based weather prediction (AIWP) models have emerged over the past 2 years, mostly in the private sector. There is an urgent need to evaluate these models from a meteorological perspective, but access to the output of these models is limited. We detail two new resources to facilitate access to AIWP model output data in the hope of accelerating the investigation of AIWP models by the meteorological community. First, a 3-yr (and growing) reforecast archive beginning in October 2020 containing twice daily 10-day forecasts forFourCastNet v2-small,Pangu-Weather, andGraphCast Operationalis now available via an Amazon Simple Storage Service (S3) bucket through NOAA’s Open Data Dissemination (NODD) program (https://noaa-oar-mlwp-data.s3.amazonaws.com/index.html). This reforecast archive was initialized with both the NOAA’s Global Forecast System (GFS) and ECMWF’s Integrated Forecasting System (IFS) initial conditions in the hope that users can begin to perform the feature-based verification of impactful meteorological phenomena. Second, real-time output for these three models is visualized on our web page (https://aiweather.cira.colostate.edu) along with output from the GFS and the IFS. This allows users to easily compare output between each AIWP model and traditional, physics-based models with the goal of familiarizing users with the characteristics of AIWP models and determine whether the output aligns with expectations, is physically consistent and reasonable, and/or is trustworthy. We view these two efforts as a first step toward evaluating whether these new AIWP tools have a place in forecast operations. 

Abstract Vertical profiles of temperature and dewpoint are useful in predicting deep convection that leads to severe weather, which threatens property and lives. Currently, forecasters rely on observations from radiosonde launches and numerical weather prediction (NWP) models. Radiosonde observations are, however, temporally and spatially sparse, and NWP models contain inherent errors that influence short-term predictions of high impact events. This work explores using machine learning (ML) to postprocess NWP model forecasts, combining them with satellite data to improve vertical profiles of temperature and dewpoint. We focus on different ML architectures, loss functions, and input features to optimize predictions. Because we are predicting vertical profiles at 256 levels in the atmosphere, this work provides a unique perspective at using ML for 1D tasks. Compared to baseline profiles from the Rapid Refresh (RAP), ML predictions offer the largest improvement for dewpoint, particularly in the middle and upper atmosphere. Temperature improvements are modest, but CAPE values are improved by up to 40%. Feature importance analyses indicate that the ML models are primarily improving incoming RAP biases. While additional model and satellite data offer some improvement to the predictions, architecture choice is more important than feature selection in fine-tuning the results. Our proposed deep residual U-Net performs the best by leveraging spatial context from the input RAP profiles; however, the results are remarkably robust across model architecture. Further, uncertainty estimates for every level are well calibrated and can provide useful information to forecasters. 

Abstract Artificial intelligence (AI) can be used to improve performance across a wide range of Earth system prediction tasks. As with any application of AI, it is important for AI to be developed in an ethical and responsible manner to minimize bias and other effects. In this work, we extend our previous work demonstrating how AI can go wrong with weather and climate applications by presenting a categorization of bias for AI in the Earth sciences. This categorization can assist AI developers to identify potential biases that can affect their model throughout the AI development life cycle. We highlight examples from a variety of Earth system prediction tasks of each category of bias. 

This project developed a pre-interview survey, interview protocols, and materials for conducting interviews with expert users to better understand how they assess and make use decisions about new AI/ML guidance. Weather forecasters access and synthesize myriad sources of information when forecasting for high-impact, severe weather events. In recent years, artificial intelligence (AI) techniques have increasingly been used to produce new guidance tools with the goal of aiding weather forecasting, including for severe weather. For this study, we leveraged these advances to explore how National Weather Service (NWS) forecasters perceive the use of new AI guidance for forecasting severe hail and storm mode. We also specifically examine which guidance features are important for how forecasters assess the trustworthiness of new AI guidance. To this aim, we conducted online, structured interviews with NWS forecasters from across the Eastern, Central, and Southern Regions. The interviews covered the forecasters’ approaches and challenges for forecasting severe weather, perceptions of AI and its use in forecasting, and reactions to one of two experimental (i.e., non-operational) AI severe weather guidance: probability of severe hail or probability of storm mode. During the interview, the forecasters went through a self-guided review of different sets of information about the development (spin-up information, AI model technique, training of AI model, input information) and performance (verification metrics, interactive output, output comparison to operational guidance) of the presented guidance. The forecasters then assessed how the information influenced their perception of how trustworthy the guidance was and whether or not they would consider using it for forecasting. This project includes the pre-interview survey, survey data, interview protocols, and accompanying information boards used for the interviews. There is one set of interview materials in which AI/ML are mentioned throughout and another set where AI/ML were only mentioned at the end of the interviews. We did this to better understand how the label “AI/ML” did or did not affect how interviewees responded to interview questions and reviewed the information board. We also leverage think aloud methods with the information board, the instructions for which are included in the interview protocols.