Using a neural network approach and time series data from an international monitoring station in the Yellow Sea for modeling marine ecosystems

Abstract

The international marine ecological safety monitoring demonstration station in the Yellow Sea was developed as a collaborative project between China and Russia. It is a nonprofit technical workstation designed as a facility for marine scientific research for public welfare. By undertaking long-term monitoring of the marine environment and automatic data collection, this station will provide valuable information for marine ecological protection and disaster prevention and reduction. The results of some initial research by scientists at the research station into predictive modeling of marine ecological environments and early warning are described in this paper. Marine ecological processes are influenced by many factors including hydrological and meteorological conditions, biological factors, and human activities. Consequently, it is very difficult to incorporate all these influences and their interactions in a deterministic or analysis model. A prediction model integrating a time series prediction approach with neural network nonlinear modeling is proposed for marine ecological parameters. The model explores the natural fluctuations in marine ecological parameters by learning from the latest observed data automatically, and then predicting future values of the parameter. The model is updated in a "rolling" fashion with new observed data from the monitoring station. Prediction experiments results showed that the neural network prediction model based on time series data is effective for marine ecological prediction and can be used for the development of early warning systems.[PUBLICATION ABSTRACT]

Environ Monit Assess (2014) 186:515524 DOI 10.1007/s10661-013-3396-8

Using a neural network approach and time series data from an international monitoring station in the Yellow Sea for modeling marine ecosystems

Yingying Zhang & Juncheng Wang & A. M. Vorontsov &

Guangli Hou & M. N. Nikanorova & Hongliang Wang

Received: 18 July 2012 /Accepted: 23 August 2013 /Published online: 21 September 2013 # Springer Science+Business Media Dordrecht 2013

Abstract The international marine ecological safety monitoring demonstration station in the Yellow Sea was developed as a collaborative project between China and Russia. It is a nonprofit technical workstation designed as a facility for marine scientific research for public welfare. By undertaking long-term monitoring of the marine environment and automatic data collection, this station will provide valuable information for marine ecological protection and disaster prevention and reduction. The results of some initial research by scientists at the research station into predictive modeling of marine ecological environments and early warning are described in this paper. Marine ecological processes are influenced by many factors including hydrological and meteorological conditions, biological factors, and human activities. Consequently, it is very difficult to incorporate all these influences and their interactions in a deterministic or analysis model. A prediction model integrating a time series prediction approach with neural network nonlinear

modeling is proposed for marine ecological parameters. The model explores the natural fluctuations in marine ecological parameters by learning from the latest ob-served data automatically, and then predicting future values of the parameter. The model is updated in a rolling fashion with new observed data from the monitoring station. Prediction experiments results showed that the neural network prediction model based on time series data is effective for marine ecological prediction and can be used for the development of early warning systems.

Keywords Marine ecological environment . Monitoring . Prediction . Time series . Neural network

Introduction

In recent years, threats such as marine pollution, overfishing, and climate change have led to the degradation of large areas of the marine environment. The decline of marine environments and the need for marine ecological protection has increasingly received worldwide attention. International cooperation will be essential to confront many of the threats to marine environments, and international collaborations have become a mainstream trend in the marine development, protection, and science research.

China has more than 18,000 km of coastline and 3,000,000 km2 of sea areas under its jurisdiction. Marine ecosystems are highly biodiverse and are also critical for the sustainable development of marine-dependent economies in both coastal and island regions. Automated long-

Y. Zhang : J. Wang (*) : G. Hou : H. WangShandong Provincial Key Laboratory of Ocean Environment Monitoring Technology, Shandong Academy of Sciences Institute of Oceanographic Instrumentation, No 28, Zhejiang Road, Qingdao 266001, Chinae-mail: [email protected]

J. Wange-mail: [email protected]

A. M. Vorontsov : M. N. NikanorovaBaltic Ecological Policy and Regulation Institute, No 4-6, 17th Liniya, Vasilievsky Ostrov,St. Petersburg 199034, Russia

516 Environ Monit Assess (2014) 186:515524

term monitoring, in conjunction with the collection of marine environmental data in real time, can help us understand the marine ecological environment and potential forecast and provide early warnings of some marine ecological emergencies. Development of these monitoring systems will be very important in marine ecological protection.

In 2009, scientists from China and Russia initiated a partnership to build an international marine ecological safety monitoring demonstration station designed as a base for monitoring the marine ecological environment and carrying out research in the Yellow Sea. In addition to scientific research, this nonprofit workstation has also been designed as a platform for public welfare which will be used for marine environment protection and marine disaster prevention and reduction. It currently provides scientists all over the world with opportunities to collaborate on marine ecological research. It is hoped that this workstation and its achievements in scientific research would not only be enjoyed by participating countries, but also be used to serve all humanity.

This paper provides a brief introduction to the design and construction of the international marine ecological safety monitoring demonstration station in the Yellow Sea. The station has now been operating for more than half a year and a substantial amount of marine ecological and environmental data has been collected and saved for scientific analysis. Some numerical predictions and early warning research studies on the marine ecological environment have been conducted. A detailed description of the typical ecological prediction is provided in the rest of the paper.

International marine ecological safety monitoring demonstration station in the Yellow Sea

The international marine ecological safety monitoring demonstration station in the Yellow Sea is based in Qingdao. It consists of a shoreside testing station, a maritime buoyage station, and an onshore data center (Fig. 1). The shoreside station is made up of marine

Fig. 1 Shoreside station and maritime buoyage station

Environ Monit Assess (2014) 186:515524 517

ecological testing laboratories and an environmental testing equipment center. These furnish the capacity for scientists to undertake laboratory-based research and to commission instruments for diverse experiments at the shoreside station. The maritime buoyage station consists of a number of disk-shaped moored buoys which are 10 m in diameter which have been designed to carry a wide variety of marine monitoring instruments. These include hydrological, meteorological, and water quality monitoring instruments. These instruments gather data, which are then transmitted wirelessly to the shore-based data center in real time. The data center is responsible for receiving, displaying, storing, and processing these data. It has a large information storage capacity and is also a platform for data analysis and marine ecological research. At present, one main research area is the evaluation of the marine ecological environment. Another equally important research area is the development of marine ecological prediction capacities to improve our ability to provide early warnings of marine ecological events, mainly algal blooms.

Measuring parameters and technical indicators

The maritime buoyage station is responsible for long-term automated marine monitoring in the Yellow Sea. Some maritime buoyage stations are reasonably deployed and then make up the marine ecological environment monitoring network.

The equipment carried by the maritime buoyage station is determined by the requirements of scientists and includes instruments to measure marine hydrological, meteorological, and water quality characteristics. A number of conventional and widely-used monitoring devices have been used on the buoys. In addition, water quality analyzers jointly developed by scientists from China and Russia have also been carried. Table 1 presents all of the parameters that can currently be measured by the buoyage system.

Marine water quality analyzers jointly developed by Russia and China

Conventional marine water quality analysis processes, collecting at sea and transporting back to lab on land

Table 1 Parameters currently measured by the maritime buoyage system

No Measuring parameter Measuring range Measuring accuracy Measuring resolution

1 Wind speed 060 m/s 0.3 m/s 0.1 m/s

2 Wind direction 0359 3 1

3 Air temperature 2050 C 0.2 C 0.1 C

4 Air pressure 8001,100 hPa 0.3 hPa 0.1 hPa

5 Relative humidity 0100 % 5 % 1 %

6 Azimuth 0359 5 1

7 Surface temperature 5 to 40 C 0.15 C 0.01 C

8 Wave Wave height: 0.225 m Wave height: 0.1 m, +5 % Wave height: 0.1 m

Wave period: 230 s Wave period: 0.25 s Wave period: 0.1 s

Wave direction: 0359 Wave direction: 10 % Wave direction: 0.5

9 Salinity 070 1 % 0.01

10 Turbidity 01,000 NTU 2 % or 0.3 NTU 0.1 NTU

11 pH 014 0.2 0.01

12 Chlorophyll 0400 g/L 5 % 0.1 g/L13 Dissolved oxygen 050 mg/L 020 mg/L: 1 % or 0.1 mg/L 0.01 mg/L 2050 mg/L: 15 %

14 Ammonia nitrogen 0700 g/L 3 % 1 g/L15 Phosphate 0400 g/L 3 % 1 g/L16 COD 010 mg/L 10 % 5 %

17 TOC 0.0110 mg/L-c 0.1 (1) 10 % (>1)

518 Environ Monit Assess (2014) 186:515524

for analysis, have a number of limitations. They often involve complex procedures, take a long time, and do not produce results representative of seawater sampling in the field. In addition, secondary pollution may be produced. Since 2003, marine water quality analyzers have been jointly developed by scientists from China and Russia, which are able to measure water quality parameters on the spot, continuously and in real-time. These instruments have been used many times for monitoring enteromorpha in the Yellow Sea and in other ocean ecological environmental investigations. At the international marine ecological safety monitoring demonstration station in the Yellow Sea, two kinds of water quality analyzers have been jointly developed and have been carried by the maritime buoyage system. These are total organic carbon (TOC) and chemical oxygen demand (COD) analyzers. Water sampling and distribution systems are also installed in the buoyage system to help TOC and COD analyzers work. The primary aim of this research was to provide us with information about TOC and COD in the seawater. In addition, it provides a further application of jointly developed water quality analyzers for long-term and real-time marine environmental monitoring.

COD analyzer is based on the principle that different pollutants generate different luminous intensities when oxidized by ozone (Liu et al. 2008). No additional reagents are needed and no secondary pollution is produced. In addition, its measurement accuracy is not affected by the concentration of chloridion in the sea-water. It can be used for automatic, continuous, and spot measurements of the chemical oxygen demand in the seawater. Its detection limit is 0.1 mg and its response speed is less than 5 min. The power consumption of this instrument is less than 200 W.

The TOC analyzer can be used for fast and spot measurements of total organic carbon in the seawater. It is based on the principles of ozone oxidation chemiluminescence dynamics. It calculates total organic carbon by using the integral of the time series on the chemical luminescence dynamics curve (Liu et al. 2011). Measurement resolution and accuracy are increased with the oxidation capacity and chemiluminescence efficiency. The analyzer had a response speed of less than 5 min.

Data acquisition and processing

At the international marine ecological safety monitoring demonstration station in the Yellow Sea, a universal data

acquisition device has been designed and developed to maximize compatibility and reliability. It offers great flexibility for users and means that equipment can be added or deleted according to need in different parts of the ocean. It has been used to greatly improve data acquisition, system control, and information transmission (Zhang et al. 2011). The maritime buoyage system stores observed data in a CompactFlash memory card and simultaneously transmits them wirelessly to the shore-based data center wirelessly. Currently, a CDMA (Code Division Multiple Access) communication method is used. Other communication methods as GPRS (General Packet Radio Service) and Beidou and Iridium Satellites are used stand-by methods.

The maritime buoyage system has the flexibility to set the time period of data acquisition and transmission according to the user demand. Currently, the system collects data on ammonia, nitrogen, phosphate, COD, and TOC two times a day. Data are collected on all other parameters once every 10 min.

Research into marine ecological prediction

The international marine ecological safety monitoring demonstration station generates a large continuous dataset, which can be used to increase our understanding of the marine ecological environment. Scientists have been using these data to conduct research into marine ecological prediction, early warning, and evaluation. This information can be used by governments to inform protection, planning, and management of marine environments. In the following section, we present the first results of our work into marine ecological prediction.

Problem analysis

To carry out research into predictive marine ecological modeling, early warning, and evaluation, it is necessary to understand the physical, meteorological, chemical, hydrological, and biological factors in the open ocean as well as near human activities. Marine ecosystems have a complicated relationship with these factors and behave in a random, uncertain, and nonlinear manner (Chou et al. 2011; Yang et al. 2008). The exact nature of the relationships between the various factors and dynamic change processes in the system is currently unknown. This has restricted the capacity of human society to carry out effective marine research in this area.

Environ Monit Assess (2014) 186:515524 519

With the development of mathematical theory and computing power, modeling and simulation techniques have provided new ways to increase our ecological understanding. A number of deterministic models have been developed which are based on a number of assumptions (Andreas and Gunther 2003; Kannel et al. 2007; Los et al. 2008; Xu and Hood 2006). Other analysis models are based on multivariate statistical methods (Arhonditsis et al. 2007; Camdevyren et al. 2005; Chau and Muttil 2007; Chou et al. 2011; Shrestha and Kazama 2007). Finally, there are also other models termed uncertainty models (Chen and Mynett 2006; Malmaeus et al. 2003). All of these types of models have their respective advantages and disadvantages as incorporating the complexities of marine ecosystem processes into any of these models is extremely challenging. The ever-increasing volume, diversity, and complexity of marine environmental data available have meant that artificial intelligence methods are increasingly used for ecological and environmental research. A general survey of relevant documents reveals that most ecological modeling and prediction research has been carried out on reservoir, river, lake, and surface water ecosystems (Kuo et al. 2007; Kusiak et al. 2012; Yao et al. 2007). Relatively few results have been published in the field of marine ecological prediction. The general regression neural networks (GRNNs) were used to predict the dissolved oxygen (DO) and chlorophyll-a (Chl-a) based on the ecological variables weekly measured between January 1997 and May 1997 in three stations of the East Johor Strait (Palani et al. 2008). Using location (longitude and latitude), temperature, salinity, and DO at lagged time from stations 1 and 3 as inputs, the DO prediction model accurately simulated the range of DO values at station 2. Using location, temperature, phosphate (PO4), DO, and

Chl-a at lagged time from stations 1 and 3, GRNN architecture performed best for the prediction of Chl-a at station 2. The adaptive filters were also used to predict seawater ecological indicators based on measurements produced by the Andromeda network, a network of sensors plunged into the Thermaikos Gulf that collects aquatic numeric data concerning seawater (Hatzikos et al. 2009). The current and two previous values of pH and turbidity measurements were used to predict the pH value for the next day. The current and two previous values of pH and dissolved oxygen were used to predict the oxygen value for the next day. In the turbidity prediction, both previous values of turbidity and pH were used. The results indicated that if the measurements

remain reasonably stationary, it is possible to make 1-day-ahead predictions.

The international marine ecological safety monitoring demonstration station in the Yellow Sea has been designing to collect the continuous marine ecological data. This makes it an ideal source of information for the numerical research on the marine ecological environment. Using large amounts of measurements, the neural network is proposed to predict marine ecological indicators based on time series of themselves in the paper. It does not require a preknowledge of the correlations and other variables.

Research method

Prediction principles and description of the model

Integrating a time series approach with a neural network approach provides a useful basis for predictive modeling of marine ecological systems. In statistics, signal processing, econometrics, and mathematical finance, a time series is a sequence of data points, measured typically at successive time instants spaced at uniform time intervals. Time series analysis comprises methods for analyzing time series data to extract meaningful statistics and other characteristics of the data (George and Jenkins 1976). Time series prediction is the use of a model to predict future values based on previously observed values. Marine ecological processes are influenced by many dynamic factors such as hydrology, meteorology, biological factors, and human activities. As a result, it is very difficult to describe all these influences in a deterministic or analysis model. However, by using time series prediction techniques, historical data can help us to describe dynamic marine ecological environments. We have adopted a promising nonlinear modeling method which incorporates these principals where an artificial neural network is developed which learns from historical marine ecological data.

The neural network prediction model based on time series data is designed according to the structure in Fig. 2. The backpropagation (BP) algorithm used in the model is that most widely used for training artificial neural networks. It is a multilayer map function that uses error backpropagation, which is able to form an associative memory and then describes an unknown system after learning from its input and output data. Because of its strong nonlinear mapping ability, this BP network has been used not only to model the causal

520 Environ Monit Assess (2014) 186:515524

x(i)

x(i-1)

x(i-n-1)

x(i+1)

Input Layer Hidden Layer Output Layer

Fig. 2 Structure of the neural network prediction model based on time series data

relationship between multiple parameters, but also to model single parameters which show temporal autocorrelation. This approach is particularly appropriate for the marine ecological parameter learning and prediction. In our research, a three-layer feedforward BP network is used. i is the sampling instant. n input neurons of the network are x(i), x(i1)x(in1), which represent n continuous observed data before the sampling instant of i. The hidden layer has m neurons. The output layer has one neuron and uses a linear transfer function. The error function in the network training is the average variance.

One of the problems associated with using the standard BP algorithm is that it has a slow convergence rate and frequently falls into local optimum points. To overcome this, a modified BP algorithm, the Levenberg Marquardt algorithm is used for network training. This modified algorithm gives better convergence rates and avoids falling into local minimum points by regulating the damping parameter according to the sign symbol and value of the average variance. It has been proposed to approximate approach Hessian matrix by iteration so to save the computational time. Weights are updated according to the following equation:

W JT J I

1JTe 1

Where J is the Jacobian of error e to weight W, and the relationship with Hessian matrix can be described as H=JTJ. is a scalar for the adaptive regulation. In the training process, the dynamic regulation of ensures the strictly decreasing of error in every step.

The Jacobi matrix is calculated more easily than the Hessian matrix based on the error backpropagation.

Then, the descending gradient of the error can be described as g=JTe.

Sample selection and pretreatment

Data from the international marine ecological safety monitoring demonstration station in the Yellow Sea, collected in 2012 by the maritime buoyage station near Eight Gap Qingdao port, were used for experiments designed to predict marine ecological parameters. Neural network prediction methods were used to model individual parameters using continuous data from the monitoring station as the sample dataset. First, these data were then separated into training and test sets in chronological order. Then, the original observed data were normalized. The data from the international marine ecological safety monitoring demonstration station were rescaled by a minmax normalization method in the range [0, 1], and then used as the sample dataset for the neural network. This kind of sample pretreatment improves training speed and also prediction accuracy.

Network construction and parameters setting

A three-layer network is used in our research as shown in Fig. 2. Its input, hidden, and output layers have n, m, and 1 neurons, respectively. The hyperbolic tangent S-type (sigmoid) transfer function is used as the active function. LevenbergMarquardt algorithm is used for network training. Using the MATLAB neural network toolkit, the weights are initialized with data automatically. Other important parameter setting is described below.

First, the number of input neurons, n, is a key parameter in the development of a prediction model. The length of historical data used for each prediction step is very important. The neural network prediction model uses the time series data to extract meaningful statistics and other characteristics of the historical ob-served data to predict future values of the marine ecological parameters. If n is too little, it is difficult to extract meaningful statistics and other characteristics of the observed data. If n is too big, in addition to undesirable effects of increasing model complexity and the length of time required for training, data may be included in the model that is irrelevant to current

Environ Monit Assess (2014) 186:515524 521

processes of interest. Tests showed that the number of input neurons needs to be decided in the prediction experiment on the different ecological parameter or the specific sea region.

Second, the number of hidden neurons, m, is also important. Using too few neurons in the hidden layers will result in the underfitting, which means that the network is unable to adequately detect the parameters in a complicated dataset. Using too many neurons can increase the time it takes to train the network and even result in overfitting. The approximate range of m is determined in our work based on three rule-of-thumb methods such as the following:

m < n1 2

m <

p n 1 a 3

m log2n 4

where a is a constant between 0 and 10. The correct m can be acquired through the gradual increase or decrease according to the certain experiment results.

Third, various more-or-less heuristic arguments have been put forward for the best choice for the damping parameter . In our experiments, the initial value of the damping parameter is 0=0.01. The scale factor is =10. Initially setting =0 and computing the error function after one step from the starting point with the damping factor of =0 and secondly with 0/. If both of these are worse than the initial point, then the damping is increased by successive multiplication by until a better point is found with a new damping factor of 0k for some k.

Last, the final aim of our marine ecological prediction model was to help us explore the dynamic nature of marine environments and to examine how they vary over time. This information will be essential for the development of early warning systems and marine ecological protection. Unlike in the industrial predictive control, the prediction accuracy of the marine ecological parameter is not emphasized. Considering the training rate, the objective mean-square deviation during the training process of neural network is set to 0.08.

Prediction experiments and results

For each different ecological parameter or specific sea region, the optimal setting of sample size and some adjustable parameters in the prediction model such as the number of input neurons and hidden neurons all needed to be tested in experiments. The following experiments and results of Chl-a and turbidity are described in detail.

1. Prediction experiments for Chl-aThe optimum sample size was determined as

830 through Chl-a prediction experiments based on the observed data collected near Eight Gap Qingdao port. The training sample size was 800 and test sample size was 30. The numbers of input neurons and hidden neurons were 30 and 68, respectively. Chl-a observations data from 24 March 2012 (01:20) to 29 March 2012 (18:10) were selected as the training sample. The learning process of the Chl-a network prediction model is shown in Fig. 3a. It may be observed that model converges rapidly. Figure 3b shows the training results, which illustrate that the network model learned the historical Chl-a data very effectively. Chl-a data from29 March 2012 (18:20) to 29 March 2012 (23:10) were used as the test sample. Test results are shown in Fig. 3c. The predicted data generally fitted the actual observed data well. The mean relative error of prediction results was 0.02909 and the squared error was 0.05906. To better illustrate the models effectiveness and to simulate the real prediction, the sample data were updated in a rolling fashion on one step based on the previous experiment. Chl-a observations from 24 March 2012 (01:30) to 29 March 2012 (18:20) were used as training sample for the prediction model. Chl-a data from 29 March 2012 (18:30) to 29 March 2012 (23:20) were used as the test sample. Keeping all the model parameters unchanged, the prediction results are shown in Fig. 3d. The model still predicted the Chl-a well. The mean relative error and squared error of prediction results were 0.01506 and 0.04196, respectively. The same rolling prediction experiments have been done for many times and all have satisfied results.

522 Environ Monit Assess (2014) 186:515524

2. Prediction experiments for turbidityThe optimum sample size of turbidity network

model was determined as 583 through prediction experiments based on observed data near Eight Gap Qingdao port. The training sample size was 553 and test sample size was 30. The optimum number of input neurons was 29 and that of hidden neurons was 70. Turbidity observation data from 1 January 2012 (00:00) to 4 January 2012 (20:00) were used as the training sample for the turbidity prediction model. Figure 4a shows the learning process of the model which converged quickly. Training results are shown in Fig. 4b. The network model learned the historical turbidity data very well. In the next step, turbidity data from 4 January 2012 (20:10) to 5 January 2012 (01:00) were used as the test sample for the turbidity prediction model. Test results are shown in Fig. 4c. Again, the predicted data were generally a good fit to the actual observed data. The mean relative error and squared error of prediction results were 0.00922 and 0.04545, respectively. Similar to the experiments on Chl-a, the sample data of the turbidity prediction model were updated in a rolling fashion on one step based on the previous experiment. Turbidity data from 1 January 2012 (00:10) to4 January 2012 (20:10) were used as the training sample and data from 4 January 2012 (20:20) to5 January 2012 (01:10) as test sample. The results of this prediction model when all the model parameters are unchanged are shown in Fig. 4d. The predictive data still showed good agreement with the actual observed data. The mean relative error was 0.00668 and squared error was0.01913. The rolling prediction experiments were carried out many times, and all satisfactorily fit the observed data.

Research results

Many prediction experiments on different marine ecological parameters have been completed based on the data from the international marine ecological safety monitoring demonstration station in the Yellow Sea in 2012. Above are the results of prediction experiments on Chl-a and turbidity. The experiments results showed that a neural network prediction model based on time series data can be used to explore the dynamics of marine

a

b

3

observed data output predicted data

2.5

Chlorophyll(g/L)

2

1.5

1

0.5

0 0 100 200 300 400 500 600 700 800

Network output data point

c

d

Fig. 3 Experiment results of the chlorophyll prediction model

Environ Monit Assess (2014) 186:515524 523

ecological parameters by learning the latest observed data, and then predicting their future values. The prediction results are satisfactory for marine ecological prediction and early warning. More importantly, a major strength of our network prediction model was that it was constantly automatically updated through constant learning from the latest observed data.

The marine ecological environment is affected by a lot of different factors. It is difficult to describe these factors and influences clearly in the models. The results of this experiment show that scientists can extract meaningful statistics and other characteristics of marine ecological parameters from their historical observed data and use them to understand their varying patterns. It has also demonstrated the value of building the international marine ecological safety monitoring demonstration station in the Yellow Sea, and of collecting substantial amounts of continuous marine ecological data. Meanwhile, we are encouraged to continue research into predictive modeling of marine environments, early warning, and evaluation using a multiparameter fusion algorithm.

Conclusions

Long-term monitoring, automatic data collection, and integrated research on prediction, early warning, and evaluation of the marine ecological environment are of major significance for scientific understanding and environment protection. The international marine ecological safety monitoring demonstration station in the Yellow Sea, an international collaboration of scientists from China and Russia, has been built as a research platform for marine ecological monitoring and analysis. At present, the demonstration station is taking shape and has been operating continuously for more than 6 months. In the future, more observation stations will be set-up and the number of observation instruments carried will be increased. Increased capacity will mean that more observational data will be gathered as required in the future. We anticipate that sustained research into marine ecological predictive modeling will be conducted and that results will be used to inform ecosystem management and for the development of early warning systems.

Acknowledgments This study was financially supported by International Scientific & Technical Cooperation Program of China (no. 2009DFB20610)

a

5

observed data output predicted data

b

4

Turbidity(NTU)

3

2

1

0 0 100 200 300 400 500 600

Network output data point

c

d

Fig. 4 Experiment results of the turbidity prediction model

524 Environ Monit Assess (2014) 186:515524

References

Andreas, M., & Gunther, R. (2003). Review of three-dimensional ecological modeling related to the North Sea shelf system. Part 1: models and their results. Progress in Oceanography, 57(2), 175217.

Arhonditsis, G. B., Paerl, H. W., Valdes-Weaver, L. M., Stow, C.A., Steinberg, L. J., & Reckhow, K. H. (2007). Application of Bayesian structural equation modeling for examining phytoplankton dynamics in the Neuse River Estuary (North Carolina, USA). Estuarine, Coastal and Shelf Science, 72, 6380.

Camdevyren, H., Demyr, N., Kanik, A., & Keskyn, S. (2005). Use of principal component scores in multiple linear regression models for prediction of chlorophyll-a in reservoirs. Ecological Modelling, 181(4), 581589.

Chau, K., & Muttil, N. (2007). Data mining and multivariate statistical analysis for ecological system in coastal waters. Journal of Hydroinformatics, 9(4), 305317.

Chen, Q. W., & Mynett, A. E. (2006). Modelling algal blooms in the Dutch coastal waters by integrated numerical and fuzzy cellular automata approaches. Ecological Modelling, 199(1), 7381.

Chou, W. R., Fang, L. S., Wang, W. H., & Tew, K. S. (2011). Environmental influence on coastal phytoplankton and zoo-plankton diversity: a multivariate statistical model analysis. Environmental Monitoring and Assessment. doi:http://dx.doi.org/10.1007/s10661-011-2373-3

Web End =10.1007/ http://dx.doi.org/10.1007/s10661-011-2373-3

Web End =s10661-011-2373-3 .

George, E. P. B., & Jenkins, G. M. (1976). Time series analysis: forecasting and control (revised edition). San Francisco: Holden Day.

Hatzikos, E., Htnen, J., Bassiliades, N., Vlahavas, I., & Fournou, E.

(2009). Applying adaptive prediction to sea-water quality measurements. Expert Systems with Applications, 36, 67736779. Kannel, P. R., Lee, S., Kanel, S. R., Lee, Y. S., & Ahn, K. (2007).

Application of QUAL2Kw for water quality modeling and dissolved oxygen control in the river Bagmati. Environmental Monitoring and Assessment, 125(13), 201217.

Kuo, J. T., Hsieh, M. H., Lung, W. S., & She, N. (2007). Using artificial neural network for reservoir eutrophication predication. Ecological Modelling, 200(12), 171177.

Kusiak, A., Verma, A., & Wei, X. P. (2012). A data-mining approach to predict influent quality. Environmental Monitoring and Assessment. doi:http://dx.doi.org/10.1007/s10661-012-2701-2

Web End =10.1007/s10661-012-2701-2 .

Liu, Y., Bai, Q., Hou, G. L., Du, L. B., & Chen, J. L. (2008). Quick analysis of seawater COD by ozone oxidation chemiluminescence. Marine Environmental Science, 27(2), 182 185 (in Chinese).

Liu, Y., Ren, G. X., & Wang, J. T. (2011). Research and implementation of total organic carbon of seawater using the chemiluminescence dynamic method. International Conference on Control, Automation and Systems Engineering, 3, 152155.

Los, F., Villars, M., & Van, M. (2008). A 3-dimensional primary production model (BLOOM/GEM) and its applications to the (southern) North Sea (coupled physical chemical ecological model). Journal of Marine Systems, 74(12), 259294.

Malmaeus, J. M., Malmaeus, J. M., & Hakanson, L. (2003). A dynamic model to predict suspended particulate matter in lakes. Ecological Modelling, 167(3), 247262.

Palani, S., Liong, S. Y., & Tkalich, P. (2008). An ANN application for water quality forecasting. Marine Pollution Bulletin, 56, 15861597.

Shrestha, S., & Kazama, F. (2007). Assessment of surface water quality multivariate statistical techniques: a case study of the Fuji river basin, Japan. Environmental Modelling and Software, 22, 464475.

Xu, J., & Hood, R. (2006). Modeling biogeochemical cycles in Chesapeake Bay with a coupled physicalbiological model. Estuarine, Coastal and Shelf Science, 69(12), 1946.

Yang, X. E., Wu, X., Hao, H. L., & He, Z. L. (2008).

Mechanisms and assessment of water eutrophication. Journal of Zhejiang University. Science. B, 9(3), 197209. Yao, Z. H., Fei, M. R., Li, K., Kong, H. N., & Zhao, B. (2007).

Recognition of blue-green algae in lakes using distributive genetic algorithm based neural networks. Neurocomputing, 70(46), 641647.

Zhang, Y. Y., Zhang, Y., Ma, R., Cheng, Y., & Hou, G. L. (2011). Design of ocean surveillance multifunctional data acquisition processor. International Conference on Intelligent Control and Information Technology, 4, 2326.