1. Aggregated daily runoff values from model output into monthly sums for the period 1991-2000.
2. For each basin, calculate the average monthly gridded runoff.
3. Multiply the average runoff sum for each basin by its drainage area, converting units to cfs.
4. Obtain observed monthly average streamflow (also in cfs) for each basin.
5. Calculate simple linear regression slope and intercept between observed flow and forecasted flow. For this, we are using monthly values from M_i, M_i-1, and M_i+1. As an example, if we're determining the regression for January, we will use monthly values from all Decembers, Januaries, and Februaries. This results in 30 values per linear regression. The R² value is also calculated for each regression and the basins plotted on the map are colored based on their R² value for the chosen month.

Basin averages of monthly runoff sums (kg/m2) overlaid on top of gridded sum for January 1991 using NOAH32.

Example scatterplot of observed vs. forecasted flows for HCDN site 02045500 for the January regression.

Additionally, summary statistics describing the skill of the forecasts over the period 1991-2010 are available on the forecasts tab. When a basin is selected, a graph and a table will appear beneath the map. The graph indicates the observed and forecasted streamflow values for the particular basin and chosen month for each year in 1991-2010. This is done for each month because the regression equations are determined independently for each month. The graphs also show the forecast’s prediction interval (http://pubs.usgs.gov/twri/twri4a3/pdf/chapter9.pdf) at the 95% confidence level.

Graph showing the observed and forecasted February streamflows for the Tar River at Tarboro, NC streamgauge (02083500) located in the Lower Tar River Basin. The R2 value for February at this site is 0.53.

Table of summary statistics for Tar River at Tarboro.

Beneath the graph is a table that further summarizes the skill for each year for a particular month’s forecasts. This table indicates whether the forecasted tercile and the observed tercile matched, and colors the cell green if they do. If they do not match, the tercile that the observed flow fell into is shaded blue while the tercile that the forecasted value fell into is shaded yellow. As an example, in the table below, which is also for the Tar River at Tarboro streamgauge (02083500), in 1994 both the observed and forecasted values were in the greater than 67th percentile range. In 1991, however, the observed value was less than the 33rd percentile value while the forecasted value for that month was in the >67th percentile.

How were these basins selected?
The majority of the basins that have streamflow forecasts are HCDN basins. These are unmanaged sites that typically have longterm records. Basins were further selected based on having more than one forecasted grid point within their boundaries. Following selection of these basins, several more basins were included based on user engagement.

Reservoir Forecasts

1. Forecast Portal Overview

The Forecast tab of the Integrated Water Portal provides monthly/seasonal inflows and storage forecasts for various reservoir systems throughout the Southeast US. Inflow and storage forecasts can be viewed by selecting the appropriate Forecast Type under the Reservoir Forecast Menu. Two types of forecasts are available:

• Individual Year Forecast: provides the forecasts for a single year selected by the user.
• Retrospective Forecast: users can view the forecasts for a range of available years. A skill summary is also provided for the selected years using the statistics in Section 4.

2. Inflow Forecast

All inflow forecasting model used in the portal have three main components: Predictors, Forecasting Model and Predictands. The predictors are inputs used to drive the forecasting model and can include forecasted data such as precipitation forecast (obtained from General Circulation Models) and observed data such as observed streamflow. The forecasting model can be statistical model or physically based hydrological models. The predictands are the output from the model or the variable of interest such as streamflow. Currently, inflow forecasts can be obtained using two methodologies: Principal Component Regression (PCR) and Climatology.

2.1 Climatology

The forecast provided using climatology is simply the long term monthly and seasonal average of the inflow for the reservoir of interest. For climatology every year will have the same results since we are averaging the monthly and seasonal values for all the years to obtain the climatology inflows. The inflows can be viewed at monthly or seasonal time scale as ensembles. The ensembles are displayed using a box plot but can be downloaded in text format from the results page. If a user is viewing the inflows under Retrospective Forecasts tab, all the years should display the same results with the same skills summary since climatology will not change from year to year but only from month to month or season to season. The reason we included climatology inflow is to allow a comparison to forecasted inflows from PCR or any other model. If the model forecasted inflows are better than the climatology it is better to use the model forecasts. If climatology outperforms the model during any month or season, we can conclude that the model does not have good skill during that particular month or season.

2.2 Principal Component Regression (PCR)

Statistical Downscaling of ECHAM4.5 precipitation forecasts

The approach we employ for developing streamflow forecasts is principal components regression, which is one of the commonly employed methodologies for statistical downscaling [Sankarasubramanian et al., 2008]. Given that the precipitation fields obtained from GCMs and SSTs are spatially correlated, application of Principal Components Analysis (PCA) rotates the original GCM fields into orthogonal components with the first mode representing the maximum variance of the original GCM fields. PCA, also known as empirical orthogonal function (EOF) analysis, on the predictors (GCM and SST fields) could also be performed by singular value decomposition on the spatial correlation matrix or covariance matrix of the predictors. Since PCA is scale dependent, loadings (eigenvectors or EOF patterns) obtained from the covariance matrix and the correlation matrix are different. Importance of each EOF pattern is quantified by the fraction of the variance the principal component represents with reference to the original predictor variance. In the case of PCA performed using correlation matrix approach, the sum of all eigenvalues is equal to the total number of elements in the data.

Dimension Reduction and Predictor Selection

For developing streamflow forecasts for various basins, we consider two candidate predictors for statistical downscaling: (a) Precipitation forecasts from ECHAM4.5 forced with constructed analogue SST forecasts over the Southeast US (23-40N;92-73W) (b) observed monthly streamflow at the site. The grid points of precipitation forecasts that correlate well with the observed streamflow are identified based on Spearman rank correlation (Table 1). These grid points of precipitation forecasts together with observed streamflow provide the predictor set for predicting streamflow for each basin. Since these predictors are correlated, it is better to employ principal components regression than multiple linear regression.

3. Storage Forecast

Along with inflow forecasts, the user can also view storage forecast for a particular reservoir. The inflows forecast obtained by using the models in Section 2 are used to obtain the storage forecast. The storage forecasts are developed using a water balance model as represented by the following equation:

Current monthly storage (S_t) is obtained based on previous month storage (S_t-1). The net inflows (q_t) are added to the storage while the evaporation (E_t) and releases (R_t) are subtracted. The net inflows (q_t) are obtained from the models in Section 2 as ensembles. The user may select observed releases or provide releases for multiple users (represented as i in the equation) of a reservoir system.

4. Forecast Skill Evaluation

The performance of the forecast is also summarized by comparing the conditional mean of the forecast with the observed flows using three different measures: (a) Correlation (b) Relative root mean square error (R-RMSE) (c) Mean Square Skill Score (MSSS).

Correlation:
The correlation is computed as Pearson correlation between the observed and the predicted flows (conditional mean) from the PCR. A good forecast is expected to have a correlation around 1. However, for the correlation to be statistically significant, the computed correlation should be greater than 1.96/(n-3)^0.5 where 'n' denotes the number of years of data used for computing correlation.

Relative - Root Mean Square Error:
Relative-RMSE denotes the average error in the conditional mean of the forecasts compared to the observed flows. A good forecast is expected to have R-RMSE closer to zero. Another way to summarize the performance of the mean of the forecast is using MSSS.

Mean Squared Error Skill Score:
MSSS is similar to R-RMSE but it compares the mean square error of the candidate forecast with the mean square error of the climatological forecast (which is just the mean monthly/seasonal streamflow). For a good forecast, we would expect MSSS should be closer to one. A forecast is considered to be poorer than climatology if MSSS is lesser than zero.

Rank Probabilistic Skill Score:
RPSS computes the cumulative squared error between the categorical forecast probabilities and the observed category in relevance to a reference forecast (Wilks, 1995). Low RPS indicates high skill and vice versa. Similarly, if RPSS is positive, then the forecast skill exceeds that of the climatological probabilities. For a good forecast, we would like the RPSS to be closer to one. RPSS could give an overly pessimistic view of the performance of the forecasts and it is a tough metric for evaluating probabilistic forecasts. For a detailed example on how to compute RPS and RPSS for given forecast, see Goddard et al., [2003].

5. References

Goddard, L., A.G. Barnston, and S.J.Mason, Evaluation of the IRI's "net assessment" seasonal climate forecasts 1997-2001. Bulletin of the American Meteorological Society 2003;84(12):1761-+. (.pdf)

Sankarasubramanian, A., U.Lall, and S.Espuneva, Role of Retrospective Forecasts of GCM Forced with Persisted SST anomalies in Operational Streamflow Forecasts Development, Journal of Hydrometeorology,9(2), 212-227, 2008.

Wilks, D.S., Statistical Methods in the Atmospheric Sciences, Academic Press, 1995.

This page is designed to assist users in finding sites that record data for a select parameter. For example, say you are interested in obtaining a list of sites having 100% of 7-day Average Streamflow observations between October 1, 2015 and December 31, 2015.

1. Under the "Data Selection" tab, select "Daily" from the dropdown menu. Then, under "Measuring" select "7-Day Averaged Streamflow."
2. Using the two calendar boxes, select October 1, 2015 in the first, and December 31, 2015 in the second.
3. Under "Data Restrictions" check the box that says "At least 90% of records." Then, select "100%" from the dropdown menu.
4. Click the "Update Map" button. Within a few seconds the map should update to show stations having 7-Day Averaged Streamflow observations.
5. To download a list of all the sites that meet your data requirements, click on the hamburger menu to the upper right of the map. Select the final option from the list that appears that says "Station Data (CSV)" (this is under the header "Download Map Data").
6. You can also see how many sites have been returned (in this case 437 total).
7. Beneath the map is a series of tabs that are color-coded based on the agency that is responsible for collecting data. In the example shown, all sites are maintained by USGS. Checking or unchecking the boxes beside each agency name will toggle its sites on or off on the map.
8. Lastly, you can view additional metadata about a site by hovering over it on the interactive map. Clicking on a site will display even more information in a box beneath the map.

The observed data page allows user to view current and historical station observation in addition to other useful information, such as gridded drought indices and the US drought monitor. Some key elements of this page are listed below and shown in the image on the right.

1. Under the "Date and Time" menu you can select a date of interest by clicking on the calendar to the right of the text box displaying the date. You can also select the checkbox beneath this date to view the most recently available data. When this box is checked, each site's most recent observation, from within the last week, will be plotted on the map. Below this you will see a "Station Data Observation Interval" dropwdown menu. Depending on the agency and specific parameter, observations may be available at a daily or sub-daily interval. Note that the choice of parameters (#2) will change based on whether you have selected "Daily" or a sub-daily time.
2. Under the station data dropdown menu is a list of all possible parameters that can be displayed on the map. These parameters are limited to mostly surface- and near-surface quantitative water data, but other parameters, such as precipitation, can also be selected. Upon selecting a parameter, the map will automatically update with the desired observations. Depending on the parameter, this may take anywhere from less than a second to up to a minute. To turn off the station data, re-select the first option in this dropdown menu: "Pick a Parameter."
   1. Once site data has been plotted, you can interact with the data by hovering over sites to see the observation. By clicking on a site, you can view a time series of observations.
   2. On the graph, you can zoom vertically by clicking within the plot and dragging. To zoom horizontally, use the bar beneath the graph.
   3. Depending on the variable, climatological percentiles information may be plotted. You can toggle these lines on or off by clicking their legend element to the right of the graph. The selected parameter's observations can also be toggle on or off by selecting it from the legend.
   4. The graph can be downloaded in several format using the hamburger menu to the upper right of the graph. From this, you can select to download an image of the graph in it's current view as a PNG or JPEG image, as a vector image in PDF or SVG format, or the data in the graph as a CSV or Excel Table document. You can also view an HTML table of the data by selecting the final option on this list; the table will plot beneath the graph.

3. Next on the left menu is Gridded Data. Beside this title is a dropdown menu with seventeen different possible timescales ranging from one day to 36 months. With the exception of KBDI (the Keetch-Byram Drought Index), all the gridded data options are associated with a time scale. For example, you can choose to view 30-day or 60-day Percent of Normal Precipitation. Short descriptions of each gridded layer are below:
   • SPI: The Standardized Precipitation Index. This is a precipitation-based drought index that represents the amount of precipitation over a timescale of interest as standard deviations above or below the historical mean. The SPI shown here is generated using AHPS precipitation estimates with historical distributions determined using historical NWS COOP station data and 1981-2010 PRISM normals. SPI is updated daily for time scales ranging from 30-days to 36-months.
   • SPI Blend: The Blended Standardized Precipitation Index. This is identical to the SPI, except that the precipitation accumulations are weighted, both for the observed amounts as well as when calculating the historical distributions. The motivation for weighting the precipitation is the assumption that precipitation that fell more recently will have a greater influence on the current level of dryness or wetness. SPI, in contrast, weights all precipitation over a given time scale equally, so rain that fell 30 days ago will have just as much weight as rain that fell yesterday. SPI Blend is updated daily for time scales ranging from 30-days to 36-months.
   • SPEI (beta): The Standardized Precipitation Evapotranspiration Index is the newest addition to the Water Portal. Computationally similar to SPI, the SPEI is based on the climate balance between precipitation (P) and potential evapotranspiration (PET). Like the SPI, SPEI values represent standard deviations above or below the historical mean P minus PET value for the given timescale. Because SPEI incorporates a temperature component, it can identify droughts that are exacerbated by heat. SPEI is updated daily for time scales ranging from 30-days to 36-months. Daily PET is estimated using daily temperatures from PRISM. Occasionally there are delays in the availability of this data. In these instances, daily RTMA temperature data is utilized until the daily PRISM estimates become available.
   • Percent of Normal Precipitation: This is the amount of precipitation over a given timescale divided by the normal amount for the same period, then multiplied by 100 to yield a percent. Daily AHPS precipitation estimates are used to calculate the observed precipitation amount. Monthly 1981-2010 PRISM normals are divided by the number of days in a month, then multiplied by the number of days needed to fill each period to calculate moving-window normal precipitation amounts. Percent of Normal Precipitation is updated daily for every available timescale from 1 day to 36 months.
   • Accumulated Precipitation: This is the amount of precipitation, in inches, over a given timescale calculated by summing daily AHPS precipitation estimates Accumulated Precipitation is updated daily for every available timescale from 1 day to 36 months.
   • KBDI: The Keetch-Byram Drought Index is the only drought index with a fixed timescale. KBDI is typically used by foresters to assess the climatological potential for fire. KBDI values are typically lowest in the winter and highest in the summer, Values are based on the previous day's value and are adjusted up or down depending on the current day's temperature and rainfall amounts. KBDI has no timescale, so adjusting the dropdown menu for timescale will not adjust the KBDI map displayed.
   • Departure from Normal Precipitation: This is the amount of precipitation over a given timescale minus by the normal amount for the same period, in inches. Daily AHPS precipitation estimates are used to calculate the observed precipitation amount. Monthly 1981-2010 PRISM normals are divided by the number of days in a month, then multiplied by the number of days needed to fill each period to calculate moving-window normal precipitation amounts. Percent of Normal Precipitation is updated daily for every available timescale from 1 day to 36 months.
4. The hamburger menu to the upper right of the map can be used to download map information. The First three options allow you to download the map, with all current data overlays as well as the legends to the right of the map, as a PNG, JPEG, or PDF image. A list of all the plotted stations and their observations can be downloaded as a CSV file. If a gridded layer is displayed on the map, its data can be downloaded as a geoTIFF in geographic (latitude/longitude) coordinates.
5. Under the Miscellaneous menu, US Drought Monitor maps can be viewed either as a shaded map or as outlines. The USDM map date is pre-determined to be the map for the closest date to your selected date.
6. Under the "Map Options" menu you can select a different region of the map to zoom into. You can also zoom and pan on the map using your mouse, or by using the "+" and "-" buttons in the upper-right of the map area. Beneath the dropdown menu for jumping to different areas of the map are options for overlaying a different base layer. You can select between a topographic map (default), a map with streets information, or a plain map. Below this is a menu for optionally displaying additional geographic or political layers to assist with interpreting data on the map. The possible options are states (displayed by default), counties, HUCs (2, 4, 6, 8, 10, or 12-digit), Lakes and Rivers, or NEXRAD Radar Coverage. You can also check the box beside "Counties" to show the county names. When this is checked, you can hover your mouse over the map and a popup box will appear with the counties name.

Monthly and seasonal forecasts for a variety of basins have been developed using output from NASA LIS. Additionally, gridded monthly forecasts of soil moisture percentiles have been generated using NASA LIS. These can be viewed under the "Gridded Data" menu. To view a time series of forecasted output:

1. Select the month-day of interest, for example December 2015.
2. Then select "Forecast Time Series" under the Basin Forecasts Menu. The map should automatically update with new basins. These are color-coded based on their R² value (5; see "Methodology" tab).
3. Click on the basin you would like to see a forecast for. In the example, the Deep River at Moncure, NC basin has been selected.
4. Below the map you will see a time series chart display with the forecasted and observed flows. You can click the series in the legend to the right of the graph to show/hide each. Additionally, you can click the hamburger menu to the upper right of the chart to export it as an image or download its data.

Under the "Reservoir Forecast Menu" (not shown) are seasonal and monthly inflow and storage forecasts for two reservoirs: Falls Lake and Jordan Lake. More detailed information about these forecasts can be found under the "Methodology: Reservoir Forecasts" tab.