Statistical and Hybrid Methods Implemented in a Web Application for Predicting Reservoir Inflows during Flood Events

Reservoir management is a critical component of flood management, and information on reservoir inflows is particularly essential for reservoir managers to make real‐time decisions given that flood conditions change rapidly. This study's objective is to build real‐time data‐driven services that enable managers to rapidly estimate reservoir inflows from available data and models. We have tested the services using a case study of the Texas flooding events in the Lower Colorado River Basin in November 2014 and May 2015, which involved a sudden switch from drought to flooding. We have constructed two prediction models: a statistical model for flow prediction and a hybrid statistical and physics‐based model that estimates errors in the flow predictions from a physics‐based model. The study demonstrates that the statistical flow prediction model can be automated and provides acceptably accurate short‐term forecasts. However, for longer term prediction (2 h or more), the hybrid model fits the observations more closely than the purely statistical or physics‐based prediction models alone. Both the flow and hybrid prediction models have been published as Web services through Microsoft's Azure Machine Learning (AzureML) service and are accessible through a browser‐based Web application, enabling ease of use by both technical and nontechnical personnel.     

Examining Modeling Approaches for the Rainfall‐Runoff Process in Wildfire‐Affected Watersheds: Using San Dimas Experimental Forest

Wildfire can significantly change watershed hydrological processes resulting in increased risks for flooding, erosion, and debris flow. The goal of this study was to evaluate the predictive capability of hydrological models in estimating post‐fire runoff using data from the San Dimas Experimental Forest (SDEF), San Dimas, California. Four methods were chosen representing different types of post‐fire runoff prediction methods, including a Rule of Thumb, Modified Rational Method (MODRAT), HEC‐HMS Curve Number, and KINematic Runoff and EROSion Model 2 (KINEROS2). Results showed that simple, empirical peak flow models performed acceptably if calibrated correctly. However, these models do not reflect hydrological mechanisms and may not be applicable for predictions outside the area where they were calibrated. For pre‐fire conditions, the Curve Number approach implemented in HEC‐HMS provided more accurate results than KINEROS2, whereas for post‐fire conditions, the opposite was observed. Such a trend may imply fundamental changes from pre‐ to post‐fire hydrology. Analysis suggests that the runoff generation mechanism in the watershed may have temporarily changed due to fire effects from saturation‐excess runoff or subsurface storm dominated complex mechanisms to an infiltration‐excess dominated mechanism. Infiltration modeling using the Hydrus‐1D model supports this inference. Results of this study indicate that physically‐based approaches may better reflect this trend and have the potential to provide consistent and satisfactory prediction.     

Regression modeling of streamflow, baseflow, and runoff using geographic information systems

Regression models for predicting total streamflow (TSF), baseflow (TBF), and storm runoff (TRO) are needed for water resource planning and management. This study used 54 streams with >20 years of streamflow gaging station records during the period October 1971 to September 2001 in Pennsylvania and partitioned TSF into TBF and TRO. TBF was considered a surrogate of groundwater recharge for basins. Regression models for predicting basin-wide TSF, TBF, and TRO were developed under three scenarios that varied in regression variables used for model development. Regression variables representing basin geomorphological, geological, soil, and climatic characteristics were estimated using geographic information systems. All regression models for TSF, TBF, and TRO had R(2) values >0.94 and reasonable prediction errors. The two best TSF models developed under scenarios 1 and 2 had similar absolute prediction errors. The same was true for the two best TBF models. Therefore, any one of the two best TSF and TBF models could be used for respective flow prediction depending on variable availability. The TRO model developed under scenario 1 had smaller absolute prediction errors than that developed under scenario 2. Simplified Area-alone models developed under scenario 3 might be used when variables for using best models are not available, but had lower R(2) values and higher or more variable prediction errors than the best models.     

COMPARATIVE EVALUATION OF STATISTICAL METHODS FOR WATER SUPPLY FORECASTING1

ABSTRACT: While the correlation coefficient and standard error of estimate are frequently used when comparing models of seasonal water yield, the following criteria may be more important in selecting one model from among several alternatives: rationality of the regression coefficients, the distribution of the residual errors, and the correctness of indicators of the relative importance of the predictor variables. These criteria were used to compare seasonal water yield models that were calibrated using multiple regression, stepwise regression, principal components regression, polynomial regression using a principal components rotation, and constrained pattern search. Hydrologic data from the Upper Sevier River basin in southern Utah were used to illustrate the comparative analysis process. The prediction equations used the April-July streamflow volume as the criterion variable.     

A Comparison of Statistical Approaches for Predicting Stream Temperatures Across Heterogeneous Landscapes1

Abstract: Estimating stream temperatures across broad spatial extents is important for regional conservation of running waters. Although statistical models can be useful in this endeavor, little information exists to aid in the selection of a particular statistical approach. Our objective was to compare the accuracy of ordinary least‐squares multiple linear regression, generalized additive modeling, ordinary kriging, and linear mixed modeling (LMM) using July mean stream temperatures in Michigan and Wisconsin. Although LMM using low‐rank thin‐plate smoothing splines to measure the spatial autocorrelation in stream temperatures was the most accurate modeling approach; overall, there were only slight differences in prediction accuracy among the evaluated approaches. This suggests that managers and researchers can select a stream temperature modeling approach that meets their level of expertise without sacrificing substantial amounts of prediction accuracy. The most accurate models for Michigan and Wisconsin had root mean square errors of 2.0‐2.3°C, suggesting that only relatively coarse predictions can be produced from landscape‐based statistical models at regional scales. Explaining substantially more variability in stream temperatures likely will require the collection of finer‐scale hydrologic and physiographic data, which may be cost prohibitive for monitoring and assessing stream temperatures at regional scales.     

Predicting Debris Yield From Burned Watersheds: Comparison of Statistical and Artificial Neural Network Models1

Abstract: Alluvial fans in southern California are continuously being developed for residential, industrial, commercial, and agricultural purposes. Development and alteration of alluvial fans often require consideration of mud and debris flows from burned mountain watersheds. Accurate prediction of sediment (hyper‐concentrated sediment or debris) yield is essential for the design, operation, and maintenance of debris basins to safeguard properly the general population. This paper presents results based on a statistical model and Artificial Neural Network (ANN) models. The models predict sediment yield caused by storms following wildfire events in burned mountainous watersheds. Both sediment yield prediction models have been developed for use in relatively small watersheds (50‐800 ha) in the greater Los Angeles area. The statistical model was developed using multiple regression analysis on sediment yield data collected from 1938 to 1983. Following the multiple regression analysis, a method for multi‐sequence sediment yield prediction under burned watershed conditions was developed. The statistical model was then calibrated based on 17 years of sediment yield, fire, and precipitation data collected between 1984 and 2000. The present study also evaluated ANN models created to predict the sediment yields. The training of the ANN models utilized single storm event data generated for the 17‐year period between 1984 and 2000 as the training input data. Training patterns and neural network architectures were varied to further study the ANN performance. Results from these models were compared with the available field data obtained from several debris basins within Los Angeles County. Both predictive models were then applied for hind‐casting the sediment prediction of several post 2000 events. Both the statistical and ANN models yield remarkably consistent results when compared with the measured field data. The results show that these models are very useful tools for predicting sediment yield sequences. The results can be used for scheduling cleanout operation of debris basins. It can be of great help in the planning of emergency response for burned areas to minimize the damage to properties and lives.     

THE EFFECT OF DATA INDEPENDENCE IN MODEL CALIBRATION AND MODEL TESTING1

ABSTRACT: With the increased use of models in hydrologic design, there is an immediate need for a comprehensive comparison of hydrologic models, especially those intended for use at ungaged locations (i.e., where measured data are either not available or inadequate for model calibration). But some past comparisons of hydrologic models have used the same data base for both calibration and testing of the different models or implied that the results of model calibration are indicative of the accuracy at ungaged locations. This practice was examined using both the regression equation approach to peak discharge estimation and a unit hydrograph model that was intended for use in urban areas. The results suggested that the lack of data independence in the calibration and testing of regression equations may lead to both biased results and misleading statements about prediction accuracy. Additionally, although split-sample testing is recognized as desirable, the split-samples should be selected using a systematic-random sampling scheme, rather than random sampling, because random sampling with small samples may lead to a testing sample that is not representative of the population. A systematic-random sampling technique should lead to more valid conclusions about model reliability. For models like a unit hydrograph model, which are more complex and for which calibration is a more involved process, data independence is not as critical because the data fitting error variation is not as dominant as the error variation due to the calibration process and the inability of the model structure to conform with data variability.     

Repeatability,Sensitivity, and Uncertainty Analyses of the BANCS Model Developed to Predict Annual Streambank Erosion Rates

Accelerated streambank erosion caused by channel instability can be the leading cause of sediment impairment of streams. Obtaining accurate streambank erosion rates for sediment budgeting and prioritizing mitigation efforts can be difficult and costly. One approach to quantifying streambank erosion rates is through the development and implementation of an empirically derived “Bank Assessment for Non‐point Source Consequences of Sediment” (BANCS) model. This study aims to improve the BANCS model application by evaluating repeatability between users and identifying sensitive and/or uncertain model inputs. Statistical analysis of streambank evaluations conducted by 10 different individuals suggests the implementation of the BANCS model may not be repeatable. This finding may be due to sensitive model inputs, such as streambank height and near‐bank stress level prediction method selection, and/or uncertain model inputs, such as bank material identification and the associated adjustment of erosion potential. Furthermore, it was found assessing streambanks as a group by obtaining a measure of central tendency from individual evaluations, as well as obtaining a higher level of training, may improve model implementation precision. Application of these suggestions may result in improved prediction of streambank erosion rates utilizing the BANCS model methodology.     

CALCULATION OF NONLINEAR CONFIDENCE AND PREDICTION INTERVALS FOR GROUND-WATER FLOW MODELS1

ABSTRACT: A method is derived to efficiently compute nonlinear confidence and prediction intervals on any function of parameters derived as output from a mathematical model of a physical system. The method is applied to the problem of obtaining confidence and prediction intervals for manually-calibrated ground-water flow models. To obtain confidence and prediction intervals resulting from uncertainties in parameters, the calibrated model and information on extreme ranges and ordering of the model parameters within one or more independent groups are required. If random errors in the dependent variable are present in addition to uncertainties in parameters, then calculation of prediction intervals also requires information on the extreme range of error expected. A simple Monte Carlo method is used to compute the quantiles necessary to establish probability levels for the confidence and prediction intervals. Application of the method to a hypothetical example showed that includsion of random errors in the dependent variable in addition to uncertainties in parameters can considerably widen the prediction intervals.     

A Multilevel Model of the Impact of Farm‐Level Best Management Practices on Phosphorus Runoff1

Abstract: Multilevel or hierarchical models have been applied for a number of years in the social sciences but only relatively recently in the environmental sciences. These models can be developed in either a frequentist or Bayesian context and have similarities to other methods such as empirical Bayes analysis and random coefficients regression. In essence, multilevel models take advantage of the hierarchical structure that exists in many multivariate datasets; for example, water quality measurements may be taken from individual lakes, lakes are located in various climatic zones, lakes may be natural or man‐made, and so on. The groups, or levels, may effectively yield different responses or behaviors (e.g., nutrient load response in lakes) that often make retaining group membership more effective when developing a predictive model than when working with either all of the data together or working separately with the individuals. Here, we develop a multilevel model of the impact of farm level best management practices (BMPs) on phosphorus runoff. The result of this research is a model with parameters which vary with key practice categories and thus may be used to evaluate the effectiveness of these practices on phosphorus runoff. For example, it was found that the effect of fertilizer application rate on farm‐scale phosphorus loss is a function of the application method, the hydrologic soil group, and the land use (crop type). Further, results indicate that the most effective method for controlling fertilizer loss is through soil injection. In summary, the resultant multilevel model can be used to estimate phosphorus loss from farms and hence serve as a useful tool for BMP selection.     

Quantifying Forest Visibility with Spatial Data

We use spatial data representing transportation networks, elevation, stand height, and recreation use to construct and compare models of recreation use patterns and visibility in a forest. The recreation use pattern model depicts use frequencies along travel corridors. The visibility model quantifies visibility for all forest areas. We find that the models provide different but complementary types of information. Forest managers who are involved in scheduling harvest operations and want to address the visual concerns of forest visitors may benefit most from the visibility model. Managers who wish to know more about travel patterns or to reroute forest visitors affected by operations may benefit from the use pattern model. A combination of the two models has the highest potential for providing planning assistance in multiple-use forests. Both models may be able to enhance visual resource management (VRM) systems already in use by providing spatially explicit recreation use and visibility data.     

COMPARISON OF MODELS FOR ESTIMATING FLOOD QUANTILES IN NEW HAMPSHIRE AND VERMONT1

We develop and compare three regression models for estimating flood quantiles at ungaged stream reaches in New Hampshire and Vermont. These models emerge from systematic analysis and validation of relations between flood magnitude and six candidate predictors reflecting basin size, topography, and climate and channel size at 36 gaging stations with record lengths exceeding 20 years. Of the candidate predictors, bank full width is most highly correlated with flood magnitude and the best prediction equation is based on width. Thus channel geometry is closely related to the current hydrologic regime in spite of geologically recent glaciation and apparently non-alluvial bank materials. We also develop models that use information obtainable from maps or GIS. The best of these uses drainage area and drainage-basin elevation as predictors, but it is substantially less precise than the width-based relation. A third relation using only drainage area as a predictor is even less precise but may be useful for some purposes. No other single predictors or combinations yielded useful predictions, although some had been included in previously-established models for the region. Model comparison included examination of residuals generated by regression using one-at-a-time suppression of data points and comparison with precision obtainable with gaging records of varying lengths.     

Improving Predictive Models of In‐Stream Phosphorus Concentration Based on Nationally‐Available Spatial Data Coverages

Spatial data are playing an increasingly important role in watershed science and management. Large investments have been made by government agencies to provide nationally‐available spatial databases; however, their relevance and suitability for local watershed applications is largely unscrutinized. We investigated how goodness of fit and predictive accuracy of total phosphorus (TP) concentration models developed from nationally‐available spatial data could be improved by including local watershed‐specific data in the East Fork of the Little Miami River, Ohio, a 1,290 km² watershed. We also determined whether a spatial stream network (SSN) modeling approach improved on multiple linear regression (nonspatial) models. Goodness of fit and predictive accuracy were highest for the SSN model that included local covariates, and lowest for the nonspatial model developed from national data. Septic systems and point source TP loads were significant covariates in the local models. These local data not only improved the models but enabled a more explicit interpretation of the processes affecting TP concentrations than more generic national covariates. The results suggest SSN modeling greatly improves prediction and should be applied when using national covariates. Including local covariates further increases the accuracy of TP predictions throughout the studied watershed; such variables should be included in future national databases, particularly the locations of septic systems.     

Problems in Extinction Model Selection and Parameter Estimation

It is a vexing problem to achieve a consensus about the proper scientific way to assess population viability for habitat conservation plans. Rather than a hypothesis-testing approach, here it is proposed to select population models, estimate extinction parameters, and assess prediction uncertainty using a pragmatic, empirical Bayesian approach. The simplest usable models include the effects of population growth, r; carrying capacity, K; Allee threshold, N(A); and environmental stochasticity, v(r). Analytic predictions of expected extinction times are available for such models. Models that are more complex can be elaborated from this basis. Selection from a hierarchy of nesting population models can often be done through the evaluation of parameters. The estimation of the most important extinction parameters can be undertaken in a variety of ways. Time series can be analyzed to estimate r(d), v(r), rho, and K. Habitat models and individualistic population models may help estimate N(A) and K and demographic stochasticity. Fine-scale biogeography and climatological data may be useful in the estimation of a variety of parameters. Because it takes many years to estimate extinction parameters accurately for a given population of interest, the most efficient estimation procedures are desirable. I propose the use of prior information from an (as yet nonexistent) population biology database. The accumulation of local information through monitoring will improve our estimates allowing adaptive management. Uncertainty in the estimates will always remain, but it may be quantified by the posterior distributions. A crude example is discussed using treefrog population data. Although the motivations, beliefs, and biases of competing stakeholders will differ, a habitat conservation plan could accommodate this variation in the prior distributions. Field experience from monitoring will increasingly clear up any discrepancies between the opposing beliefs and the real ecosystem. As the world is an uncertain place and because there is no universal scientific method, there will always be controversy and surprises. The best we can do is (1) agree about our prior information, (2) agree about the strategy of model selection and parameter estimation, and (3) agree about our strategy for adaptive management. Perhaps the greatest impediment to such prior agreements for HCPs is the likely paranoia inspired by the use of unfamiliar statistical methodology. We need to train students of ecology in a more flexible and deeper understanding of statistics and philosophy of science.     

Prediction Intervals for Estimates of Site Index Based on Ecosystem Type

/ British Columbia has an ecosystem classification system that classifies sites into site series. Foresters commonly measure the productivity of these sites by their site index. In British Columbia, site index is defined as the height of a stand at breast height age 50 and is usually estimated from height-age models. Biogeoclimatic site series/site index relationships are an increasingly popular method of estimating site index in British Columbia for stands where site index cannot be reliably estimated with height-age models. The precision of the predicted site index from these relationships can be evaluated with prediction intervals. This is done for the predicted site index of a single site, a group of sites, or the areally weighted site index of a group of sites. The methodology is also useful in determining the number of sites required to meet a specified precision. These prediction intervals will assist foresters in making sound forest management decisions.KEY WORDS: Biogeoclimatic Ecosystem Classification; Precision; Prediction interval; Site index; Site series     

ESTIMATION OF SOIL MOISTURE WITH API ALGORITHMS AND MICROWAVE EMISSION1

Large area soil moisture estimations are required to describe input to cloud prediction models, rainfall distribution models, and global crop yield models. Satellite mounted microwave sensor systems that as yet can only detect moisture at the surface have been suggested as a means of acquiring large area estimates. Relations previously discovered between microwave emission at the 1.55 cm wavelength and surface moisture as represented by an antecedent precipitation index were used to provide a pseudo infiltration estimation. Infiltration estimates based on surface wetness on a daily basis were then used to calculate the soil moisture in the surface 0–23 cm of the soil by use of a modified antecedent precipitation index. Reasonably good results were obtained (R²= 0.7162) when predicted soil moisture for the surface 23 cm was compared to measured moisture. Where the technique was modified to use only an estimate of surface moisture each three days an R² value of 0.7116 resulted for the same data set. Correlations between predicted and actual soil moisture fall off rapidly for repeat observations more than three days apart. The algorithms developed in this study may be used over relatively flat agricultural lands to provide improved estimates of soil moisture to a depth greater than the depth of penetration for the sensor.     

Post-development Auditing of EIA Predictive Techniques: A Spatial Analytical Approach

Despite widespread recognition that post-development auditing has the potential to provide feedback which could improve future Environmental Impact Assessment (EIA), there remains a paucity of research which relates specifically to the evaluation of EIA predictive techniques, with even less progress in the development of audit methodologies. This paper describes a spatial analytical approach to post-development auditing that focuses upon the identification and analysis of the residual errors between the impacts predicted at a site using a particular predictive method and the actual impacts found to occur through monitoring. For three case studies, relevant impact predictions are tested (to determine the residual errors) and statistical models of the errors are developed in order to explore factors which may explain the performance of the predictive technique. The paper then considers the broader lessons and limitations that can be drawn out from the research both for auditing and EIA practice, including feedback on predictive techniques, the potential role of scoping decisions in generating errors in impact prediction, and the implications of uncertainty over future baseline conditions for auditing and impact prediction/interpretation.     

几种大气预测模式预测结果比较分析

估算模式、AERMOD模式系统、ADMS模式系统均是HJ2.2-2008《环境影响评价技术导则大气环境》中推荐的大气预测模式,为探求此3种大气预测模式预测结果的大小关系规律,选用估算模式、AERMOD模式系统、ADMS模式系统,在简单地形和复杂地形两种条件下,结合一般工业类环评项目中常见的点源、面源案例,对不同预测模式的大气预测结果进行比较分析,得出相应的规律,对环评工作中进一步预测模式的选用具有一定的参考借鉴意义。     

EFFECTS OF STATSGO AND SSURGO AS INPUTS ON SWAT MODEL'S SNOWMELT SIMULATION1

ABSTRACT: Soil data comprise a basic input of SWAT (Soil and Water Assessment Tool) for a watershed application. For watersheds where site specific soil data are unavailable, the two U.S. Department of Agriculture (USDA) soil databases, the State Soil Geographic (STATSGO) and Soil Survey Geographic (SSURGO) databases, may be the best alternatives. Although it has been noted that SWAT models using the STATSGO and SSURGO data may give different simulation results for water, sediment, and agricultural chemical yields, information is scarce on the effects of using these two databases in predicting streamflows that are predominantly generated from melting snow in spring. The objective of this study was to assess the effects of using STATSGO versus SSURGO as an input for the SWAT model's simulation of the streamflows in the upper 45 percent of the Elm River watershed in eastern North Dakota. Designating the model as SWAT‐STATSGO when the STATSGO data were used and SWAT‐SSURGO when the SSURGO data were used, SWAT‐STATSGO and SWAT‐SSURGO were separately calibrated and validated using the observed daily streamflows. The results indicated that SWAT‐SSURGO provided an overall better prediction of the discharges than SWAT‐STATSGO, although both did a good and comparable job of predicting the high streamflows. However, SWAT‐STATSGO predicted the low streamflows more accurately and had a slightly better performance during the validation period. In addition, the discrepancies between the discharges predicted by these two SWAT models tended to be larger at upstream locations than at those farther downstream within the study area.     

ACCURACY AND PRECISION OF NRCS MODELS FOR SMALL WATERSHEDS1

ABSTRACT: In the last 30 years, the National Resource Conservation Service's TR‐55 and TR‐20 models have seen a dramatic increase in use for stormwater management purposes. This paper reviews some of the data that were originally used to develop these models and tests how well the models estimate annual series peak runoff rates for the same watersheds using longer historical data record lengths. The paper also explores differences between TR‐55 and TR‐20 peak runoff rate estimates and time of concentration methods. It was found that of the 37 watersheds tested, 25 were either over‐ or under‐predicting the actual historical watershed runoff rates by more than 30 percent. The results of this study indicate that these NRCS models should not be used to model small wooded watersheds less than 20 acres. This would be especially true if the watershed consisted of an area without a clearly defined outlet channel. This study also supports the need for regulators to allow educated hydrologists to alter pre‐packaged model parameters or results more easily than is currently permitted.