3.1. Short-Term Operational Precipitation and Flow Prediction in a Data-Rich Environment
Research on the development of coupled meteorological–hydrological model components for operational hydrologic applications over a single hydrologic basin was initiated in the early 1980s with simplified precipitation prediction components, spatially-lumped conceptual hydrologic and channel-routing models, and with explicit account of the propagation and update of uncertainty (first and second moment) based on available real-time observations of precipitation and flow [
7,
8]. Performance of the operational systems was good for good-quality precipitation and flow data and for sustainable operations. The uncertainty component was a stable extended Kalman filter, suitable for continuous model dynamics, which was later implemented as part of the operational system in the US [
3,
9].
Along those lines, one of the first operational implementations of such a system was for the 3300 km
2 mountainous Panama Canal Watershed, the waters of which are used to facilitate Panama Canal shipping (
Figure 3). Toward this end, distributed precipitation predictions and quasi-distributed land-surface models were used with state estimation [
10]. The tributaries that feed the Panama Canal drain small basins that range in size from less than 100 km
2 to approximately 700 km
2. Precipitation over the Watershed is measured by an S-band radar and a dense network of automated rain gauges. The gauge data were used to bias adjust the radar data and, in the absence of radar observations, to provide estimates of the mean areal precipitation (MAP) and its uncertainty (kriging method was used) over the basins. The flow in each of the significant tributaries is measured by automated gauges at sites that have frequent updates in rating curves. Predictions are useful for a range of lead times from 1 h to 24 h.
Under this rather observation-rich environment and with rather short useful lead times, characterizing the uncertainty in flows using automated state estimators is beneficial, and statistical errors were kept significantly below the climatological error bounds due to frequent state estimator updates. Over time, the model components of this system have been upgraded, and because of the continuing hands-on training of the operational hydrologists and meteorologists of the ACP, the system remains operational, providing information useful for the management of the Panama Canal ([
11,
12,
13]).
The stochastic–dynamic formulations that are involved in this type of implementation are at an advanced level that requires graduate-level background in uncertain dynamical systems; that is, background that is rarely available in the operational environment outside of focused graduate school study. The viable option for sustainability and utility in that situation was to focus the training on the interpretation of the uncertainty output of the operational system and on providing a prerequisite basic statistics course for operational hydrologists and meteorologists [
11,
12]. In addition and importantly, the state estimator formulations in the operational system needed to be adjusted so that real-time configuration changes in the observed data sources could be handled in a reliable manner. Three illustrations of this are discussed below.
First, in several cases, the radar data was not available in real time, and there were changes in the configuration of the rainfall-observing gauge network, as some of the sensors did not report for some of the time. In these cases, a first- and second-moment adjustment of the sub-basin MAP was made to correctly account for the time-varying reporting network and provide the forecasters with consistent uncertainty information. Second, if the real-time flow observations were not available for a particular time, the state estimator simply did not proceed with the update of the second-moment properties across an observation for the state vector. After some time had passed (e.g., 6 h), a state estimator was used to estimate the state again, with new predict–update cycles starting from some time in the past (several days ago). In many cases, the flow observations were available after the fact; therefore, a better and more stable estimate of the states and their uncertainty was obtained through this continuous data-reprocessing approach. Third, the potential incidents of system “crash” also needed to be handled in such a way that the operational forecasters could make the system operational quickly. After several configuration adjustments, the one that was found most useful for the rather small basins and response times was to estimate the steady-state limit of the second moments of the state variables for each month and keep this exported. After such a crash, these limits were used to expeditiously reinstate the real-time system operations, without the need to run the system again from the beginning of the rainy season, as would otherwise be required because of the soil water content memory. With hourly updates, the states were found to quickly adjust to the real-time data.
The information from this hydrometeorological forecast system (both the mean and the variance of the products) was used operationally by the Panama Canal Authority to determine timely actions in the case of predicted flooding in tributary streams. For the Panama Canal operations, this included extracting staff to safe ground, as well as the equipment that guides ships during Canal passage, and making decisions as to the safe operation of the Canal for shipping.
3.2. Flash-Flood Operational Prediction in a Data-Sparse Environment
The characterization of uncertainty in flow products through automated means based on automated observations, as discussed in
Section 3.1, is not feasible in large areas of the world because of a lack of such observations. In particular, when the focus is on flash flooding over large regions (sometimes encompassing several countries) with high resolution, this automated state-estimation method of uncertainty characterization is not applicable. In fact, the prediction of uncertainty is valid through state estimation or through ensemble prediction [
14], but the updating of uncertainty across observations through state estimators cannot be used in such data-sparse areas.
This situation is prevalent in the implementation of flash flood guidance systems (FFGSs) worldwide [
6,
15]. This operational system for flash flood assessment and occurrence prediction serves more than 64 countries worldwide. It is based on meteorological and hydrological models and on remotely-sensed multispectral satellite and local data (radar and on-site precipitation gauge data). The first consideration for such a system was to use coupled meteorological and hydrological models, e.g., [
8,
16], to produce assessments and predictions of flash flood occurrence. Real-time applications indicated that low-quality data significantly impact the performance of such operational systems [
17].
Flash floods typically have short durations (<6 h) and small spatial scales (<200 km
2) and are the result of rainfall, land-surface cover, and soil water saturation conditions [
18] (
Figure 4). In lieu of on-site radar and rain gauge observations from dense networks, rainfall observations from satellites carry significant errors in such small scales; thus, it is important that real-time updates are engaged to best approximate the actual land-surface conditions in regions of flash flood occurrence. Additionally, typical in these implementations is the sparsity of land-surface data, including flash flood occurrence data and streamflow data. Therefore, although the lead times are rather short and the initial conditions significantly influence the prediction (in spite of large forcing uncertainty), the benefit of having automated systems to update the land-surface states (soil water content and snow water equivalent) from observations on small flash-flood scales is not realizable. A new approach was necessary.
Research was performed on the impact of errors in small-basin, real-time estimation of precipitation for basins with radar coverage and good operational density of on-site rain gauges, also considering errors in the parameters of the land-surface components. This research indicated that the uncertainty in the simulation of flow increases linearly as the logarithm of the basin area decreases, with precipitation input contributing the largest portion of flow uncertainty [
19,
20]. The flow simulation errors are about 30% for 1000 km
2 basins and increase to about 90% for 100 km
2 basins. Flow prediction, rather than simulation, errors are expected to be much higher in smaller basins, especially for forecast lead times longer than a few hours.
Fortunately, with respect to precipitation errors, operational meteorologists and hydrologists have significant experience with specific observation networks and with predictive high-resolution mesoscale models for certain areas. This experience could be used to make adjustments in real time to precipitation observations and/or predictions if the operational diagnostic and prognostic flash flood system were designed to allow for this. This was taken into consideration in the design and subsequent enhancements of the FFGS operational system.
The approach followed was to decouple of the meteorological and hydrological component models but in a way that allows for assessments of the risk of flash flooding. Toward this goal, an early warning index was determined to be the bankfull flow of the streams at the outlets of the identified small flash-flood-prone basins. Then, a link was made between the bankfull flow and the amount and duration of certain rainfall over the catchment that could cause this bankfull flow [
21,
22]. Consequently, actual or forecast mean areal rainfall of a given duration (1 h, 3 h, or 6 h) that is greater than this certain rainfall of the same duration and over the same basin would yield exceedance of the bankfull flow at the small basin and indicates likelihood of flash flooding. This certain rainfall is termed the flash flood guidance of the given duration.
In this manner, the observed or forecast rainfall becomes a product of the system, allowing for forecaster adjustments, the impact of which on the exceedance of the flash flood guidance may be directly identified by forecasters. Appropriate interactive interfaces were designed to facilitate this process. Such interfaces provided separate information for observed and forecast precipitation, surface soil water saturation, and flash flood guidance of various durations. Through these interactive interfaces, the forecaster can look at several scenarios using precipitation bounds or, in some cases, even use the interface to make several adjustments directly to precipitation and produce adjusted products so that a final adjusted product may be selected for the final prediction of flash flood occurrence. Naturally, training of operational forecasters on the basis of the system model components and the effective use of the interfaces is a critical component for sustainability, and a significant hands-on training program has been developed to support forecasters.
Forecaster adjustments to the observed rainfall are warranted based on up-to-the-minute information (observer reports, video feeds, and local gauges) not included in the current cycle of system computations and/or on the forecaster’s experience with the local reliability of gauge-corrected satellite information. These adjustments yield more accurate simulations and better initial conditions for the next forecast.
Forecaster adjustments to future rainfall are based on prior numerical weather prediction-model validation studies and the experience of forecasters with the forecasts of specific numerical weather prediction models for specific seasons and regions within their country. They tend to reduce the forecast model biases, providing better overall assessment for the likelihood of future flash flood occurrence. In some of the current implementations of this approach, the systems have been made flexible to receive input from more than one numerical weather prediction model and to show the products for each of these models separately. This allows forecasters to make a single model selection in real time or consult more than one model to make their final adjustments.
Evaluations of this forecaster-centric approach that decouples the precipitation and land-surface response components in regions with very sparse data and for trained forecasters indicates that the forecaster adjustments made in real time provide significant skill (reduction in assessment uncertainty) for the identification of flash flood events in small basins [
23]. It also allows for implementations under a variety of available data and forecasts.