FLOOD STAGE FORECASTING WITH SUPPORT VECTOR MACHINES1

ABSTRACT: Machine learning techniques are finding more and more applications in the field of forecasting. A novel regression technique, called Support Vector Machine (SVM), based on the statistical learning theory is explored in this study. SVM is based on the principle of Structural Risk Minimization as opposed to the principle of Empirical Risk Minimization espoused by conventional regression techniques. The flood data at Dhaka, Bangladesh, are used in this study to demonstrate the forecasting capabilities of SVM. The result is compared with that of Artificial Neural Network (ANN) based model for one‐lead day to seven‐lead day forecasting. The improvements in maximum predicted water level errors by SVM over ANN for four‐lead day to seven‐lead day are 9.6 cm, 22.6 cm, 4.9 cm and 15.7 cm, respectively. The result shows that the prediction accuracy of SVM is at least as good as and in some cases (particularly at higher lead days) actually better than that of ANN, yet it offers advantages over many of the limitations of ANN, for example in arriving at ANN's optimal network architecture and choosing useful training set. Thus, SVM appears to be a very promising prediction tool.     

Soil Moisture Data Assimilation Using Support Vector Machines and Ensemble Kalman Filter1

Abstract: A hybrid data assimilation (DA) methodology that combines two state‐of‐the‐art techniques, support vector machines (SVMs) and ensemble Kalman filter (EnKF), is applied for soil moisture DA in this work. The SVM methodology provides a statistically sound and robust approach to solving the inverse problem, and thus to building statistical models. EnKF is an extension of the Kalman Filter (KF), a well‐known tool in prediction updating. In the present research, ground measurements were used to build a SVM‐type soil moisture predictor. Subsequent observations and their statistics were assimilated to update predictions from the SVM model by coupling it with EnKF. In this way, both model predictions and ground data, as well as their statistics, are fused thus minimizing the prediction error and making the predictions and observations statistically consistent. The results are shown for two approaches; one in which update is done at every time step and the other which assumes that data is only available at alternate time steps (in window of 10 time steps) and hence update is performed at those occasions. The SVM‐EnKF coupling is shown to improve soil moisture forecasts in an example using data from the Soil Climate Analysis Network site at Ames, Iowa.     

Approximating SWAT Model Using Artificial Neural Network and Support Vector Machine1

Abstract: With the popularity of complex, physically based hydrologic models, the time consumed for running these models is increasing substantially. Using surrogate models to approximate the computationally intensive models is a promising method to save huge amounts of time for parameter estimation. In this study, two learning machines [Artificial Neural Network (ANN) and support vector machine (SVM)] were evaluated and compared for approximating the Soil and Water Assessment Tool (SWAT) model. These two learning machines were tested in two watersheds (Little River Experimental Watershed in Georgia and Mahatango Creek Experimental Watershed in Pennsylvania). The results show that SVM in general exhibited better generalization ability than ANN. In order to effectively and efficiently apply SVM to approximate SWAT, the effect of cross‐validation schemes, parameter dimensions, and training sample sizes on the performance of SVM was evaluated and discussed. It is suggested that 3‐fold cross‐validation is adequate for training the SVM model, and reducing the parameter dimension through determining the parameter values from field data and the sensitivity analysis is an effective means of improving the performance of SVM. As far as the training sample size, it is difficult to determine the appropriate number of samples for training SVM based on the test results obtained in this study. Simple examples were used to illustrate the potential applicability of combining the SVM model with uncertainty analysis algorithm to save efforts for parameter uncertainty of SWAT. In the future, evaluating the applicability of SVM for approximating SWAT in other watersheds and combining SVM with different parameter uncertainty analysis algorithms and evolutionary optimization algorithms deserve further research.     

Comparison of Machine Learning Models Performance on Simulating Reservoir Outflow: A Case Study of Two Reservoirs in Illinois,U.S.A.

Reservoir outflow is an important variable for understanding hydrological processes and water resource management. Natural streamflow variation, in addition to the streamflow regulation provided by dams and reservoirs, can make streamflow difficult to understand and predict. This makes them a challenge to accurately simulate hydrologic processes at a daily scale. In this study, three Machine Learning (ML) algorithms, Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN), were examined and compared to model reservoir outflow. Past, current, and future hydrologic and meteorological data were used as model inputs, and the outflow of next day was used as prediction. Simulation results demonstrated that all three models can reasonably simulate reservoir outflow. For Carlyle Lake, the coefficient of determination and Nash–Sutcliffe efficiency were each close to one for the three models. The coefficient of determination, relative mean bias, and root mean square error indicated that the SVM performed better than the RF and ANN, but the SVM output displayed a larger relative mean bias than that from RF and ANN. For Lake Shelbyville, the ANN model performed better than RF and SVM when considering the coefficient of determination, Nash–Sutcliffe efficiency, relative mean bias, and root mean square error. The study results demonstrate that the three ML algorithms (RF, SVM, and ANN) are all promising tools for simulating reservoir outflow. Both the accuracy and efficacy of the three ML algorithms are considered to support practitioners in planning reservoir management.     

IMPROVED RAINFALL/RUNOFF ESTIMATES USING REMOTELY SENSED SOIL MOISTURE1

ABSTRACT: Remotely sensed soil moisture data measured during the Southern Great Plains 1997 (SGP97) experiment in Oklahoma were used to characterize antecedent soil moisture conditions for the Soil Conservation Service (SCS) curve number method. The precipitation‐adjusted curve number and the soil moisture were strongly related (r²= 0.70). Remotely sensed soil moisture fields were used to adjust the curve numbers and the runoff estimates for five watersheds, in the Little Washita watershed; the results ranged from 2.8 km² to 601.6 km². The soil moisture data were applied at two spatial scales, a finer one (800 m) measuring spatial resolution and a coarser one (28 km). The root mean square error (RMSE) and the mean absolute error (MAE) of the runoff estimated by the standard SCS method was reduced by nearly 50 percent when the 800 m soil moisture data were used to adjust the curve number. The coarser scale soil moisture data also significantly reduced the error in the runoff predictions with 41 percent and 28 percent reductions in MAE and RMSE, respectively. The results suggest that remote sensing of soil moisture, when combined with the SCS method, can improve rainfall runoff predictions at a range of spatial scales.     

APPLICATION OF THE RHESSys MODEL TO A CALIFORNIA SEMIARID SHRUBLAND WATERSHED1

ABSTRACT: Distributed hydrologic models which link seasonal streamflow and soil moisture patterns with spatial patterns of vegetation are important tools for understanding the sensitivity of Mediterranean type ecosystems to future climate and land use change. RHESSys (Regional Hydro‐Ecologic Simulation System) is a coupled spatially distributed hydroecological model that is designed to be able to represent these feedbacks between hydrologic and vegetation carbon and nutrient cycling processes. However, RHESSys has not previously been applied to semiarid shrubland watersheds. In this study, the hydrologic submodel of RHESSys is evaluated by comparing model predictions of monthly and annual streamflow to stream gage data and by comparing RHESSys behavior to that of another hydrologic model of similar complexity, MIKESHE, for a 34 km² watershed near Santa Barbara, California. In model intercomparison, the differences in predictions of temporal patterns in streamflow, sensitivity of model predictions to calibration parameters and landscape representation, and differences in model estimates of soil moisture patterns are explored. Results from this study show that both models adequately predict seasonal patterns of streamflow response relative to observed data, but differ significantly in terms of estimates of soil moisture patterns and sensitivity of those patterns to the scale of landscape tessellation used to derive spatially distributed elements. This sensitivity has implications for implementing RHESSys as a tool to investigate interactions between hydrology and ecosystem processes.     

Prediction of global solar radiation using support vector machines

This article utilizes Support Vector Machines (SVM) for predicting global solar radiation (GSR) for Sharurha, a city in the southwest of Saudi Arabia. The SVM model was trained using measured air temperature and relative humidity. Measured data of 1812 values for the period from 1998–2002 were obtained. The measurement data of 1600 were used for training the SVM, and the remaining 212 were used for comparison between the measured and predicted values of GSR. The GSR values were predicted using the following four combinations of data sets: (i) Daily mean air temperature and day of the year as inputs, and global solar radiation as output; (ii) daily maximum air temperature and day of the year as inputs, and GSR as output; (iii) daily mean air temperature and relative humidity and day of the year as inputs, and GSR as output; and (iv) daily mean air temperature, day of the year, relative humidity, and previous day’s GSR as inputs, and GSR as output. The mean square error was found to be 0.0027, 0.0023, 0.0021, and 7.65 × 10^?4 for case (i), (ii,), (iii), and (iv) respectively, while the corresponding absolute mean percentage errors were 5.64, 5.08, 4.48, and 2.8%. Obtained results show that the SVM method is capable of predicting GSR from measured values of temperature and relative humidity.     

COMPARISON OF PROCESS‐BASED AND ARTIFICIAL NEURAL NETWORK APPROACHES FOR STREAMFLOW MODELING IN AN AGRICULTURAL WATERSHED1

ABSTRACT: The performance of the Soil and Water Assessment Tool (SWAT) and artificial neural network (ANN) models in simulating hydrologic response was assessed in an agricultural watershed in southeastern Pennsylvania. All of the performance evaluation measures including Nash‐Sutcliffe coefficient of efficiency (E) and coefficient of determination (R²) suggest that the ANN monthly predictions were closer to the observed flows than the monthly predictions from the SWAT model. More specifically, monthly streamflow E and R² were 0.54 and 0.57, respectively, for the SWAT model calibration period, and 0.71 and 0.75, respectively, for the ANN model training period. For the validation period, these values were ?0.17 and 0.34 for the SWAT and 0.43 and 0.45 for the ANN model. SWAT model performance was affected by snowmelt events during winter months and by the model's inability to adequately simulate base flows. Even though this and other studies using ANN models suggest that these models provide a viable alternative approach for hydrologic and water quality modeling, ANN models in their current form are not spatially distributed watershed modeling systems. However, considering the promising performance of the simple ANN model, this study suggests that the ANN approach warrants further development to explicitly address the spatial distribution of hydrologic/water quality processes within watersheds.     

PERCOLATE WATER AND BROMIDE MOVEMENT IN THE ROOT ZONE OF EFFLUENT IRRIGATION SITES1

ABSTRACT: A bromide tracer was used to evaluate percolate water and ion movement in the upper 1.2 m of soil at a proposed sewage effluent irrigation site located in the Missouri Ozarks. Two plots representing Doniphan silt loam and Crider silt loam soils were sprinkler irrigated with local ground water at a rate of 7.62 cm/week from June through August 1976. Soil water potential, percent soil moisture by volume, and background levels of bromide in soil water, ground water, and precipitation were measured at the study plots. Bromide exchange properties and saturated hydraulic conductivity of the soils were determined in the laboratory. During two selected time periods, irrigation water, was spiked with NaBr (5.0 mg/l Br). Bromide movement through the upper profile was quantified by soil water samples and post-sampling neutron activation analysis. Soil moisture was near saturatin in both soils when the Br tracer was applied. Bromide concentrations above background levels (0.023 mg/l Br, Doniphan silt loam and 0.016 mg/l Br, Crider silt loam) were detected within 2.60 hours at 0.9 m in the Doniphan soil and within 3.75 hours at that depth in the Crider soil. The rate of Br movement in the profile was greater in both soils than the measured saturated hydraulic conductivity, Bromide concentrations above background levels were present in soil water from the study plots for a minimum of 21 days after irrigation with the Br tracer.     

Reconstructions of Soil Moisture for the Upper Colorado River Basin Using Tree‐Ring Chronologies1

Anderson, SallyRose, Glenn Tootle, and Henri Grissino‐Mayer, 2012. Reconstructions of Soil Moisture for the Upper Colorado River Basin Using Tree‐Ring Chronologies. Journal of the American Water Resources Association (JAWRA) 48(4): 849‐858. DOI: 10.1111/j.1752‐1688.2012.00651.x Abstract: Soil moisture is an important factor in the global hydrologic cycle, but existing reconstructions of historic soil moisture are limited. We used tree‐ring chronologies to reconstruct annual soil moisture in the Upper Colorado River Basin (UCRB). Gridded soil moisture data were spatially regionalized using principal components analysis and k‐nearest neighbor techniques. We correlated moisture sensitive tree‐ring chronologies in and adjacent to the UCRB with regional soil moisture and tested the relationships for temporal stability. Chronologies that were positively correlated and stable for the calibration period were retained. We used stepwise linear regression to identify the best predictor combinations for each soil moisture region. The regressions explained 42‐78% of the variability in soil moisture data. We performed reconstructions for individual soil moisture grid cells to enhance understanding of the disparity in reconstructive skill across the regions. Reconstructions that used chronologies based on ponderosa pines (Pinus ponderosa) and pinyon pines (Pinus edulis) explained more variance in the datasets. Reconstructed soil moisture data was standardized and compared with standardized reconstructed streamflow and snow water equivalent data from the same region. Soil moisture and other hydrologic variables were highly correlated, indicating reconstructions of soil moisture in the UCRB using tree‐ring chronologies successfully represent hydrologic trends.     

Stochastic Transfer Function Model and Neural Networks to Estimate Soil Moisture1

Abstract: A practical methodology is proposed to estimate the three‐dimensional variability of soil moisture based on a stochastic transfer function model, which is an approximation of the Richard’s equation. Satellite, radar and in situ observations are the major sources of information to develop a model that represents the dynamic water content in the soil. The soil‐moisture observations were collected from 17 stations located in Puerto Rico (PR), and a sequential quadratic programming algorithm was used to estimate the parameters of the transfer function (TF) at each station. Soil texture information, terrain elevation, vegetation index, surface temperature, and accumulated rainfall for every grid cell were input into a self‐organized artificial neural network to identify similarities on terrain spatial variability and to determine the TF that best resembles the properties of a particular grid point. Soil moisture observed at 20 cm depth, soil texture, and cumulative rainfall were also used to train a feedforward artificial neural network to estimate soil moisture at 5, 10, 50, and 100 cm depth. A validation procedure was implemented to measure the horizontal and vertical estimation accuracy of soil moisture. Validation results from spatial and temporal variation of volumetric water content (vwc) showed that the proposed algorithm estimated soil moisture with a root mean squared error (RMSE) of 2.31% vwc, and the vertical profile shows a RMSE of 2.50% vwc. The algorithm estimates soil moisture in an hourly basis at 1 km spatial resolution, and up to 1 m depth, and was successfully applied under PR climate conditions.     

On the Relationship Between Mean and Variance of Soil Moisture Fields1

Abstract: The objective of this work was to explain an apparent contradiction in the literature related to the relationship between mean and variance (or standard deviation) of soil moisture fields. Some studies found an increase in soil moisture variance with decreasing mean soil moisture, while others showed a decrease. The evidence of maximum variance in the mid‐range of mean soil moisture was also reported in the literature. In this paper, we focus on the effects of spatial variability of soil texture on the relationship between variance and mean of soil moisture during soil dry‐down processes. Soil texture influences soil moisture mean and variance through its direct effects on evaporation and drainage, which are two main factors controlling soil drying. A differential equation describing soil moisture dry down is proposed and studied. Our study shows that as mean soil moisture is greater than a threshold, variance increases with decreasing mean soil moisture. If mean soil moisture is less than the threshold, variance decreases with decreasing mean soil moisture. The threshold depends on soil texture and is between the field capacity and the wilting point. The soil moisture dry‐down equation is also applied to explain the apparent contradiction with regard to the relationship between mean and variance of soil moisture fields reported in the literature.     

Estimating Root Zone Soil Moisture at Continental Scale Using Neural Networks

This study investigates the feasibility of artificial neural networks (ANNs) to retrieve root zone soil moisture (RZSM) at the depths of 20 cm (SM20) and 50 cm (SM50) at a continental scale, using surface information. To train the ANNs to capture interactions between land surface and various climatic patterns, data of 557 stations over the continental United States were collected. A sensitivity analysis revealed that the ANNs were able to identify input variables that directly affect the water and energy balance in root zone. The data important for RZSM retrieval in a large area included soil texture, surface soil moisture, and the cumulative values of air temperature, surface soil temperature, rainfall, and snowfall. The results showed that the ANNs had high skill in retrieving SM20 with a correlation coefficient above 0.7 in most cases, but were less effective at estimating SM50. The comparison of the ANNs showed that using soil texture data improved the model performance, especially for the estimation of SM50. It was demonstrated that the ANNs had high flexibility for applications in different climatic regions. The method was used to generate RZSM in North America using Soil Moisture and Ocean Salinity (SMOS) soil moisture data, and achieved a spatial soil moisture pattern comparable to that of Global Land Data Assimilation System Noah model with comparable performance to the SMOS surface soil moisture retrievals. The models can be efficient alternatives to assimilate remote sensing soil moisture data for shallow RZSM retrieval.     

SOIL SALINITY PREDICTION USING ARTIFICIAL NEURAL NETWORKS1

ABSTRACT: This study explores the applicability of Artificial Neural Networks (ANNs) for predicting salt build‐up in the crop root zone. ANN models were developed with salinity data from field lysimeters subirrigated with brackish water. Different ANN architectures were explored by varying the number of processing elements (PEs) (from 1 to 30) for replicate data from a 0.4 m water table, 0.8 m water table, and both 0.4 and 0.8 m water table lysimeters. Different ANN models were developed by using individual replicate treatment values as well as the mean value for each treatment. For replicate data, the models with twenty, seven, and six PEs were found to be the best for the water tables at 0.4 m, 0.8 m and both water tables combined, respectively. The correlation coefficients between observed salinity and ANN predicted salinity of the test data with these models were 0.89, 0.91, and 0.89, respectively. The performance of the ANNs developed using mean salinity values of the replicates was found to be similar to those with replicate data. Not only was there agreement between observed and ANN predicted salinity values, the results clearly indicated the potential use of ANN models for predicting salt build‐up in soil profile at a specific site.     

Sensitivity of SCS Models to Curve Number Variation1

ABSTRACT: The Soil Conservation Service (SCS) models, including the TR-20 computer program and the simplified methods in TR-55, are widely used in hydrologic design. The runoff curve number (CN), which is an important input parameter to SCS models, is defined in terms of land use tretments, hydrologic, condition, antecedent soil moisture, and soil type. The objective of this study was to evaluate the sensitivity of the SCS models to errors in CN estimates. The results show that the effects of CN variation decrease as the design rainfall depth increases, such as for the larger storm events. The value and use of the sensitivity curves are demonstrated using a comparison of Landsat and conventionally derived curve numbers for three watersheds in Pennsylvania.     

Support vector machine based prediction of photovoltaic module and power station parameters

ABSTRACT

The uncertainty in the output power of the photovoltaic (PV) power generation station due to variation in meteorological parameters is of serious concern. An accurate output power prediction of a PV system helps in better design and planning. The present study is carried out for the prediction of output power of PV generating station by using Support Vector Machines. Two cases are considered in the present study for prediction. Case-I deals with the prediction of PV module parameters such as V_oc, I_sh, R_s, R_sh, I_max, V_max, P_max, and case-II deals with the prediction of power generation parameters such as P_DC, P_AC, and system efficiency. Historical data of PV power station with an installed capacity of 10 MW and weather information are used as input to develop four different seasons-based SVM models for all parameters. The performance results of the models are presented in terms of Mean Relative Error (MRE) and Root Mean Square Error (RMSE). Additionally, the performance results obtained with polynomial and Radial Based Function kernel are also compared to show that which kernel has better prediction accuracy, and practicability. The result shows that the minimum average RMSE and MRE for case-I with Radial Based Function kernel are 0.034%, 0.055%, 0.002%, 1.726%, 0.044%, 0.047%, 2.342%, and 0.005%, 0.014%, 0.079%, 0.885%, 0.005%, 0.007%, 0.013%, and for case-II with poly kernel are 0.014%, 0.016%, 0.149% and 0.011%, 0.0175, 1.03%, respectively. The present study will be helpful to provide technical guidance to the prediction of the PV power System.     

Predicting Particulate Matter (PM2.5) Concentrations in the Air of Shahr‐e Ray City,Iran, by Using an Artificial Neural Network

Particulate matter (PM), along with other air pollutants, pose serious hazards to human health. The Artificial Neural Network (ANN) is a branch of artificial intelligence that has an ability to make accurate predictions. In this article, the authors describe such methods and how historical data on air quality, moisture, wind velocity, and temperature in Shahr‐e Ray City, located at the southern tip of Tehran, was used to train an ANN to provide accurate predictions of PM concentrations. The availability of such predictions can offer significant assistance to those who are working to reduce air pollution.     

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.     

SPATIAL DISTRIBUTION OF POINT SOIL MOISTURE ESTIMATES USING LANDSAT TM DATA AND FUZZY-C CLASSIFICATION1

ABSTRACT: Many hydrologic models have input data requirements that are difficult to satisfy for all but a few well-instrumented, experimental watersheds. In this study, point soil moisture in a mountain watershed with various types of vegetative cover was modeled using a generalized regression model. Information on sur-ficial characteristics of the watershed was obtained by applying fuzzy set theory to a database consisting of only satellite and a digital elevation model (DEM). The fuzzy-c algorithm separated the watershed into distinguishable classes and provided regression coefficients for each ground pixel. The regression model used the coefficients to estimate distributed soil moisture over the entire watershed. A soil moisture accounting model was used to resolve temporal differences between measurements at prototypical measurement sites and validation sites. The results were reasonably accurate for all classes in the watershed. The spatial distribution of soil moisture estimates corresponded accurately with soil moisture measurements at validation sites on the watershed. It was concluded that use of the regression model to distribute soil moisture from a specified number of points can be combined with satellite and DEM information to provide a reasonable estimation of the spatial distribution of soil moisture for a watershed.     

Comparison of Field Performance of Multiple Soil Moisture Sensors in a Semi‐Arid Rangeland1

Abstract:  Automated electronic soil moisture sensors, such as time domain reflectometry (TDR) and capacitance probes are being used extensively to monitor and measure soil moisture in a variety of scientific and land management applications. These sensors are often used for a wide range of soil moisture applications such as drought forage prediction or validation of large‐scale remote sensing instruments. The convergence of three different research projects facilitated the evaluation and comparison of three commercially available electronic soil moisture probes under field application conditions. The sensors are all installed in shallow soil profiles in a well instrumented small semi‐arid shrub covered subwatershed in Southeastern Arizona. The sensors use either a TDR or a capacitance technique; both of which indirectly measure the soil dielectric constant to determine the soil moisture content. Sensors are evaluated over a range of conditions during three seasons comparing responses to natural wetting and drying sequences and using water balance and infiltration simulation models. Each of the sensors responded to the majority of precipitation events; however, they varied greatly in response time and magnitude from each other. Measured profile soil moisture storage compared better to water balance estimates when soil moisture in deeper layers was accounted for in the calculations. No distinct or consistent trend was detected when comparing the responses from the sensors or the infiltration model to individual precipitation events. The results underscore the need to understand how the sensors respond under field application and recognize the limitations of soil moisture sensors and the factors that can affect their accuracy in predicting soil moisture in situ.