Understanding the Atlas
This atlas takes the variables commonly used to describe environmental conditions and transforms them into metrics that are easier to understand, relate to, and act on. Instead of asking the user to interpret raw model output, it expresses the data as meaningful, decision-relevant quantities, and presents them at the spatial scales where real decisions are made, from individual locations up to entire states.
It does this across the southeastern United States, spanning sixteen states from Texas and Oklahoma in the west across to Maryland and Delaware in the east. Crucially, it does not rely on a single source of information. It draws on several independent sets of observations, on dozens of model simulations (downscaling output from up to 25 different global climate models), and on several different downscaling methods, so that the same environmental conditions can be examined from many independent vantage points and the uncertainty around them made visible.
That design raises a few natural questions: what does it mean to “downscale” a model, and why is it necessary? Why use many observations instead of one? Why use many models, and why more than one downscaling method? The rest of this section answers each of these in turn, because together they explain both what the atlas contains and why it is built the way it is.
What “downscaling” means, and why it is needed
Global models simulate the Earth on a relatively coarse grid. Downscaling is the process of enhancing the information contained in that coarse output to resolve the finer detail it does not natively capture, bringing the simulated signal down to the local scales at which decisions are actually made.
The need for it comes down to a mismatch of scales. A global model might represent the land surface in blocks tens of kilometers across, treating everything inside a block as a single uniform value. At that resolution, features that strongly shape local conditions are blurred away or missed entirely.
A coarse grid cell may average a stretch of coastline together with open ocean and inland terrain, so the sharp contrast between a humid coast and a drier interior is lost. Likewise, a mountain range that drives heavy rainfall on one slope and a dry rain shadow on the other can be flattened into a single intermediate value, misrepresenting both sides. For decisions that depend on exactly these local differences, the raw coarse output is of limited use, which is why it must be downscaled.
Why multiple observations
It might seem that observations should simply be “the truth,” but the gridded observational datasets used here are themselves derived products. They are constructed by blending many underlying sources, such as ground stations, radar, and satellite measurements, and interpolating them onto a continuous grid.
Because they are derived, no single observational dataset is perfect. Different choices of input sources and interpolation methods lead to real differences between them, especially in areas with sparse station coverage or complex terrain. Using several independent observational datasets, rather than trusting any one, lets a user see where they agree (and can be trusted with more confidence) and where they diverge (and should be treated with more caution).
Why so many models
Models represent environmental conditions by resolving the physical processes that connect the land, atmosphere, and ocean. Those representations are powerful, but they are inevitably incomplete, for two main reasons.
First, no model fully represents every part of the Earth system; each makes simplifying assumptions about processes that are difficult to capture exactly. Second, models run on grids that are often coarser than the scales at which important processes actually occur.
Convective processes, such as the localized thunderstorms that produce much of the summertime rainfall in the Southeast, develop at scales far smaller than a typical model grid cell. A model cannot resolve an individual storm directly, so it has to approximate the overall effect of many such storms. These approximations are useful but imperfect, and they are a major reason different models disagree.
So models do their best, but they still contain errors, and they produce information at scales quite different from the local scales needed for decision-making. Both facts point to the same conclusion: their output needs to be downscaled, and a single model is never enough. Using a large ensemble of models lets a user characterize the uncertainty that comes from the choice of model.
Why multiple downscaling methods
Downscaling itself is not perfect either. Each method makes its own assumptions about how to translate coarse information to a finer scale, and those assumptions introduce their own differences in the result. To capture this method-based uncertainty, the atlas applies several downscaling methods rather than relying on one.
Bringing these threads together, the atlas provides information at many levels of decision-making, drawing on multiple observational datasets, many models, and several downscaling methods, so that a user can understand not only what the environmental conditions look like, but how confident to be in that picture and where the uncertainty comes from.
Spatial Level
The Spatial Level menu, on the left, lets the user choose the scale at which data is aggregated and displayed. The atlas offers several levels, ranging from the most detailed to the most coarse. The underlying datasets are produced at different native resolutions, so to compare them fairly, or to combine them with other information, they first have to be expressed on the same grid or aggregated to a common spatial unit. This places datasets of differing resolutions onto a shared footing, so that observations, models, and downscaling methods can be set side by side. Each level also corresponds to a different decision-making context for a different set of stakeholders, from water managers and local planners to state policymakers and site-specific analysis. Bringing the data to these scales makes it directly usable for the people and decisions that operate at each level.
Lat-Lon Grid (0.1°)
The finest, most detailed level. Data is shown in a grid of evenly sized cells, each about 0.1° (roughly 10 km) of latitude and longitude across, giving the most precise, location-specific view. At this 10 km grid level, the data is presented in 3D, with cell height conveying the magnitude of the selected index.
Cities
Data summarized across 100 cities, with three sub-levels: Metro Areas (broader metropolitan region), Zip Codes (individual postal zones), and City Boundaries (within a city’s official limits).
Counties
Data aggregated by county, useful for sub-state comparisons and administrative reporting.
Watersheds (HUC8)
Data grouped by drainage basin using the USGS 8-digit Hydrologic Unit Code.
States
The coarsest level. Data is summarized for each state as a whole, best for broad, region-wide comparisons.
The user can choose whichever level best matches the scale of their interest, from a single grid cell up to an entire state.
Historical Climate
The Historical Climate section organizes data into three types of indexes. Temperature and precipitation indexes are available from two sources: observations, derived from recorded measurements, and model simulations. These data can be visualized at multiple spatial scales, from the native latitude–longitude grid to counties, states, and larger regions.
Temperature Indexes
Temperature indexes are derived from daily maximum (Tmax), minimum (Tmin), and mean (Tmean) temperature data. Most indexes are available at monthly, seasonal, and annual timescales. Temperature indexes are available for the period 1980–2023.
Precipitation Indexes
Precipitation indexes are derived from daily precipitation data and include all forms of precipitation, including rain and snow. Most indexes are available at monthly, seasonal, and annual timescales. Precipitation indexes are available for the period 1980–2023.
Hazard Indexes
Climate-related hazards are currently based on the NOAA Storm Events Database. For a range of weather hazards (such as floods, tornadoes, hurricanes, severe storms, hail, winter storms, heat, and drought), we provide the frequency of events along with the economic damage and the number of fatalities associated with each hazard type. Hazard data is available only from the NOAA Storm Events Database at the county scale.
The reported damage associated with these extremes is often not very accurate. It is presented here as is, while work is in progress to bring improvements to these estimates.
Some hazard records in this database extend back to the 1950s, but the early reports are incomplete and their coverage is very sparse. It is only from 1996 onward that a large suite of hazard conditions and environments began to be reported consistently, which is why the atlas uses data from 1996 to 2025.
Data Sources
Observations come from three datasets: Daymet, Livneh, and PRISM. Each provides a record of measured climate conditions, and together they offer multiple independent views of the historical climate.
| Dataset | Resolution | Coverage period | Source |
|---|---|---|---|
| Daymet | 1 km | 1980–present | daymet.ornl.gov ↗ |
| Livneh | ~6 km (1/16°) | 1915–2018 | psl.noaa.gov ↗ |
| PRISM | 4 km | 1981–present | prism.oregonstate.edu ↗ |
Although the individual datasets begin in different years, the atlas starts its analysis in 1981, since that is the earliest year for which all three observational datasets are available, ensuring a consistent common period across them.
Daymet
Produced by Oak Ridge National Laboratory’s Distributed Active Archive Center (ORNL DAAC), Daymet provides daily gridded estimates of weather parameters (including minimum and maximum temperature and precipitation) across North America at a 1-km resolution, derived by interpolating ground-based station observations and spanning 1980 to the present. Native resolution: 1 km. daymet.ornl.gov ↗
Livneh
A long-term, hydrologically consistent gridded dataset for the conterminous United States, derived from daily temperature and precipitation observations at roughly 20,000 NOAA Cooperative Observer (COOP) stations and provided at about 6 km (1/16°) resolution. It is widely used as training data for downscaling and as validation for climate and hydrologic models, and is distributed by the NOAA Physical Sciences Laboratory (PSL). Native resolution: ~6 km (1/16°). psl.noaa.gov ↗
PRISM
Developed and maintained by the PRISM Climate Group at Oregon State University, PRISM uses the Parameter-elevation Regressions on Independent Slopes Model to generate high-resolution gridded climate estimates for the conterminous United States. By combining station data with a digital elevation model, it accounts for terrain effects such as rain shadows, temperature inversions, and coastal gradients, and is considered a benchmark for U.S. climatological mapping. Native resolution: 4 km. prism.oregonstate.edu ↗
Simulations are produced by downscaling global models from the Coupled Model Intercomparison Project Phase 6 (CMIP6). All simulation datasets are bias-corrected, and some are additionally statistically downscaled. The methods differ in how they get from a coarse global model to fine-scale, bias-corrected output: one hybrid (dynamical followed by statistical), four statistical, and two AI-based. Several of these datasets are produced in-house at Oak Ridge National Laboratory (ORNL): the hybrid (RegCM), AI-based, and DBCCA datasets. The remaining datasets come from their own external sources.
The methods do not all cover the same number of global models. RegCM-based hybrid downscaling covers 6 models, the AI-based and DBCCA methods cover 10 models each, ESDM covers 14, and LOCA and the NASA dataset each cover more than 25. The atlas provides data for a maximum of 25 models. As a result, there are at least 6 global models that have been downscaled by every method, which provides a particularly useful basis for studying both model-based and method-based uncertainty side by side.
Dynamical downscaling is computationally expensive compared with the other methods, which is why the RegCM-based hybrid dataset has a smaller ensemble of models than the statistical and AI-based datasets.
| Method | Developed / hosted by | Resolution | Models | Reference |
|---|---|---|---|---|
| RegCM (hybrid) | ICTP (model); produced at ORNL | 25 km → 4 km | 6 | ICTP/RegCM ↗ |
| DBCCA | Produced at ORNL | 4 km | 10 | Werner & Cannon, 2016 ↗ |
| AI-based (SRCNN / GNN) | Produced at ORNL | 4 km | 10 | Rastogi et al., 2025 ↗ |
| ESDM | Texas Tech University | 4 km | 14 | Hayhoe et al., 2024 ↗ |
| LOCA | Scripps Institution of Oceanography | ~6 km (1/16°) | >25 | loca.ucsd.edu ↗ |
| BCSD (NEX-GDDP) | NASA | 0.25° (~25 km) | >25 | NASA NEX-GDDP-CMIP6 ↗ |
The atlas provides data for a maximum of 25 models per method.
Hybrid (dynamical + statistical) downscaling
A two-stage approach that combines dynamical and statistical downscaling. First, a physics-based regional climate model, RegCM, dynamically downscales the global climate model (GCM) by simulating the actual atmospheric processes that shape weather and climate, producing output at a 25 km grid. RegCM solves physical equations for the atmosphere over a limited area, representing local features such as terrain, coastlines, and land cover in much greater detail than a global model can. That 25 km output is then statistically downscaled and bias-corrected to a 4 km grid, giving fine-scale climate information grounded in the underlying physics of the climate system. The RegCM model is developed and maintained by the Abdus Salam International Centre for Theoretical Physics (ICTP); this dataset is produced at ORNL. Native resolution: 25 km (dynamical), downscaled to 4 km. github.com/ICTP/RegCM ↗
Statistical downscaling
Downscales GCMs directly, using mathematical relationships between large-scale climate patterns and local conditions, learned from historical data, to refine coarse data to a finer scale (with bias correction built into the process). Four statistical methods are included. Two of these, LOCA and ESDM, were the datasets used in the Fifth National Climate Assessment (NCA5). Of the four, DBCCA is produced at ORNL, while LOCA, BCSD, and ESDM come from their own external sources (linked below).
- Double Bias-Corrected Constructed Analogs (DBCCA)An analog-based method that downscales each day by combining observed days whose large-scale patterns most closely resemble the model’s, then applies a second quantile-mapping bias correction at the fine scale to remove residual biases (such as precipitation “drizzle”) introduced by combining days. Native resolution: 4 km. Werner & Cannon, 2016 ↗
- Localized Constructed Analogs (LOCA) Used in NCA5A multiscale analog method that selects a single best-matching observed day for each location rather than averaging many analogs, producing more realistic spatial detail and better preserving extremes. It was one of the two downscaled datasets used in the Fifth National Climate Assessment (NCA5). Native resolution: ~6 km (1/16°). loca.ucsd.edu ↗
- Bias-Correction Spatial Disaggregation (BCSD) In progressA quantile-mapping approach that first bias-corrects coarse model output against observations, then spatially disaggregates it to a finer grid using observed climatological patterns. In the atlas, BCSD data is drawn from NASA’s NEX-GDDP-CMIP6 global downscaled projections. Integration of this dataset is currently in progress. Native resolution: 0.25° (~25 km). NASA NEX-GDDP-CMIP6 ↗
- Empirical Statistical Downscaling Model (ESDM) Used in NCA5 In progressA signal-decomposition approach (STAR-ESDM) that separates a daily time series into long-term trends, seasonal cycles, and residuals, bias-corrects each component, and recombines them, a flexible and computationally efficient method applicable to many variables and data sources. Along with LOCA, it was one of the two downscaled datasets used in the Fifth National Climate Assessment (NCA5). Integration of this dataset into the atlas is currently in progress. Hayhoe et al., 2024 ↗
AI-based downscaling
Applies machine-learning models, specifically Convolutional Neural Networks (CNN) and Graph Neural Networks (GNN), to learn complex patterns and generate fine-scale climate data. Here the workflow first bias-corrects the data and then downscales it. This follows a super-resolution approach: rather than running an expensive physics-based model, a super-resolution CNN (SRCNN) is trained on paired coarse and dynamically downscaled fields to emulate dynamical downscaling of daily precipitation, learning to map low-resolution input to high-resolution output. Incorporating elevation data and pre-processed inputs further improves accuracy, and once trained the network can downscale global climate model output quickly and at far lower computational cost than dynamical downscaling. These datasets are produced at ORNL. Native resolution: 4 km. Rastogi et al., 2025 (GRL) ↗
Altogether, these methods produce over 100 downscaled datasets. They can be visualized individually and compared with one another at the state, county, city, grid, or watershed level.
Why a common 10 km grid
As the table shows, the datasets are produced at different native resolutions, ranging from about 4 km up to roughly 25 km. Not every dataset is available at the finest 4 km resolution, so bringing them all to 4 km would mean inventing detail that some of them simply do not contain. Instead, the atlas brings every dataset to a common, intermediate 10 km latitude-longitude grid. This resolution is high enough to capture meaningful local detail, yet coarse enough to be genuinely representative of all the datasets involved, so that they can be compared on a fair and consistent footing.
Sectors in Focus
Beyond the core environmental data, the current version of the atlas specifically targets two sectors that are especially important to the Southeast: agriculture and transportation.
Agriculture In progress
The Southeast is one of the most agriculturally productive parts of the country, and its farms are highly sensitive to weather and environmental conditions such as heat, drought, and excess moisture. The atlas focuses on six major crops grown across the region: cotton, soybean, hay, sugarcane, corn, and wheat. Because the growth, yield, and quality of these crops depend so directly on environmental stress, understanding how conditions vary and change across the region is central to the resilience of the rural economy and the communities that depend on it.
The Agriculture view brings together two kinds of data:
Agriculture indexes. Growing degree days and heat- and moisture-based measures relevant to crops. As with the other environmental data in the atlas, these indexes are available from both observations and downscaled, bias-corrected model simulations, and can be visualized at any spatial level, from the fine Lat-Lon grid up to the state scale. The model simulation data is converted into metrics designed to capture weather-related stress on crops. For each metric, a user can examine its observed range and how well it is represented in the downscaled datasets, making it possible to see both the historical conditions crops have experienced and how faithfully the models reproduce them. Continued work combines these indexes into aggregated estimates of crop stress, and those datasets will be integrated into the atlas as they become available.
Observed agricultural data. Yield, harvested area, and production records for the six major crops. Observed agricultural data is available at the county scale only.
Transportation
The Southeast is also one of the regions most exposed to weather extremes, and its transportation infrastructure sits directly in harm’s way. Roads and bridges across the region are repeatedly stressed by hurricanes, mesoscale convective systems and other severe storms, and the flooding that often accompanies them. Because this infrastructure is critical to daily life, commerce, and emergency response, understanding where it is most exposed to these hazards is an important part of building regional resilience. The Transportation view covers two areas:
Critical Infrastructure In progress. Focused on bridges across the Southeast, covering more than 250,000 bridges. Currently the atlas visualizes different characteristics of these bridges, such as their condition ratings and annual traffic volume, with the data provided at the county level. The broader goal is to analyze bridges for their vulnerability to environmental conditions and to identify the most at-risk structures. Flooding is the leading cause of bridge failures in the U.S., acting through multiple mechanisms including scour, hydraulic loading, debris accumulation, and embankment erosion. By linking these hazards to bridge characteristics, the atlas aims to surface the bridges most vulnerable to environmental stress.
Road Safety In progress. The Road Safety analysis tool, currently under development, will allow users to understand how roadway safety conditions vary under wet weather conditions across the Southeast. By combining crash data, roadway characteristics, and weather information, the tool will help identify locations and conditions associated with increased safety risks and support more informed transportation planning and decision-making.
Analysis In progress
The Analysis section is the platform’s deep-dive workspace. After selecting a spatial level, a user clicks one of the individual regions within it (a grid cell, city, county, watershed, or state) to open a detailed analysis of that region, explored through an expanding suite of analytical views that move from a high-level signal down to the factors driving it, and quantify how confident we can be in that signal.
Comparative analysis
The analysis can also compare data across multiple locations, choosing up to 10 states, counties, or watersheds at a time. This is useful because it lets a user directly compare environmental conditions, trends, and variability between regions, seeing at a glance how one part of the Southeast differs from another and which areas stand out. By viewing many regions side by side, patterns that would be hard to spot one location at a time become clear.
Time series
View how an index evolves over time at a location, making it easy to identify the prevailing long-term behavior and year-to-year variability before digging into what drives it.
Climate stripes
Anomaly plots that show each year’s departure from a 1991–2014 reference period, rendered as a sequence of colored stripes. This gives an immediate, intuitive picture of how any index has shifted relative to its baseline climate over time.
At the state level, climate stripes can be viewed both for the state as a whole and for every county within the state. Comparing them side by side reveals, for a given year, which counties contribute most to the state-level anomalies.
Trend analysis
Examine trends at monthly, seasonal, and annual scales. By decomposing an overall trend into its monthly and seasonal components, a user can see which months or seasons are driving the larger annual signal, building understanding step by step.
Probability density functions (PDFs)
Show the full distribution of an index’s values, making it easy to compare the shape, spread, and central tendency of observations against models.
Scatter plots
Plot one variable against another to expose relationships, correlations, and outliers between indexes or between sources.
Joint probability distributions
Show how two indexes co-occur, capturing the likelihood of combined conditions: for example, hot and dry occurring together.
Sankey plots
A flow diagram that brings together all of the indexes for a location into a single picture, summarizing the prevailing environmental conditions at that place, an integrated snapshot rather than one index at a time.
Spatial trends
Visualize the spatial pattern of trends across the region for any of the indexes, making it easy to see where conditions are changing fastest and how trends vary from place to place. This view is currently implemented at the county level.
At the state level, spatial trends can be compared between observations and models to understand how comparable the two are spatially. Models tend to capture the state-average trend well, but appear to have less skill in reproducing the distribution of those trends across counties within the state.
Animation
Play an animation of the annual spatial anomalies, watching how each year departs from the baseline climate across the region over time. This view is currently implemented at the county level.
Flexible comparisons
Every view (trends, time series, and climate stripes) can be compared across data sources in any combination: observations versus observations, observations versus models, or models versus models. This lets a user understand how individual models compare with observations, how models compare with one another, and how a single model compares with its own variants produced under different downscaling methods.
Uncertainty quantification
These comparisons come together in uncertainty quantification. By setting models against observations and against each other, a user can fully understand where the spread in the data comes from, whether it arises from the choice of climate model or the choice of downscaling method. To go a step further, the analysis applies the ANOVA (analysis of variance) method to formally partition the total uncertainty and identify the largest source: model or method.
Detailed Dashboard In progress
The Detailed Dashboard brings all of the analyses together on a single dashboard, so a user can see everything they need about a location in one place rather than navigating between separate views.
It is especially useful for building presentations: a user can select whichever analyses they want (time series, climate stripes, trends, PDFs, scatter plots, and more) and assemble them onto one page, producing a single, comprehensive snapshot ready to share or present.
Chatbot In development
The Chatbot is an AI-enabled query system, currently under development, that will let users explore the atlas using plain-English questions. Instead of navigating the menus manually, users will be able to simply ask for what they are looking for, and the atlas will respond with the related visualizations and data.