Research Project:
ASSESSING A NEURAL NETWORK MODELING APPROACH FOR PREDICTING NUTRIENT LOADS IN THE MAHANTANGO WATERSHED
Location: University Park, Pennsylvania
Project Number: 1902-13000-011-08
Project Type:
Specific Cooperative Agreement
Start Date: Sep 01, 2005
End Date: Aug 31, 2007
Objective:
The objective of this cooperative research project is to determine the strengths and limitations of a neural network (NN) modeling approach for water quality assessment at a watershed scale. Among various forms of artificial intelligence (AI), artificial neural network (NN) is on the rise for applications in hydrology and soil science. The strengths and limitations of deterministic process-based models, such as the Soil and Water Assessment Tool (SWAT), are well known, and serve as a standard for relative comparison among alternative modeling approaches. The specific objectives will be to: (1) develop an NN model to predict nutrient loads (nitrogen and/or phosphorus) in the Mahantango drainage basin; and (2) compare model development requirements, data requirements, and predictive capability to our existing experience and knowledge of comparable elements of SWAT.
Approach:
Our extensive stream water quality database for the ARS experimental watershed at Klingerstown, PA (Mahantango watershed) will be fully utilized in this project for developing optimal NN model(s) and for validating the simulation results. GIS will be used in this research as a spatial data manager and analyzer. GIS will also be integrated with the NN model development and application. The popular GIS software, ArcGIS, will be used in this study because of our experience with this software, its compatibility with simulation models, and its wide use in federal, state, and local governmental agencies. Spatial distributions of those attributes important to agricultural activity and nutrient movement in the watershed will be characterized. Many of these spatial data layers have already been collected in our GIS database. Artificial NN model development is data intensive. The fact that we have an extensive long-term database, will give us a solid starting point to construct NN models. Certain data screening guidelines will be followed closely before constructing NN models to ensure integrity in the data. General recommendations for NN model development include: 1) all cause-and-effect variables are measurable and used; 2) redundant variables are minimized or eliminated; 3) the variable data is uniformly distributed with sufficient points; 4) data sampling times shows the actual system response; 5) the data is properly screened for outliners and redundancy; 6) true variability exists in the data and all data is valid; 7) model accuracy requirements do not surpass measurement accuracy and variability; and 8) model prediction will not be extrapolated in any variable dimension. After these guidelines are followed, accurate and robust NN models will be developed, which then will be used for prediction and strategic management planning. The input variables, presented in the form of GIS, may include nutrient loading, slope, USDA Curve Number, soil permeability, geology, and other factors that are available and important in determining the fate of nutrients in the watershed. We plan to use a feed-forward NN modeling approach in this study. After the training phase, a validation (¿testing¿) phase without iterations serves to validate the NN models on independent input and output data by keeping weight matrices constant. We will use 10% of collected data to do the validation. If a NN model is valid, the average root-mean-square residuals for the validation phase should not differ much from those of the calibration set. In this study, evaluation criteria for the "goodness-of-fit" of the predicted and measured data would include coefficient of determination, error sum of squares, and a modified coefficient of efficiency. Our experience in developing calibrating, validating and applying the NN model will be compared to the recent accomplishment whereby Dr. Veith compared performances of three continuous, watershed-scale models (ANSWERS-2000, SWAT, and SMDR), when applied to a small northeastern watershed, and determined that SWAT most accurately simulated observed hydrologic processes.
|
|
|
|