Estimating Water Depths Using Artificial Neural Networks

The real-time Everglades Depth Estimation Network (EDEN) has been established in the Everglades of South Florida, USA, to support a variety of scientific and water-management purposes. The expansiveness of the Everglades, limited number of gauging stations, and extreme sensitivity of fauna to small changes in water depth have created a need for accurately predicting water depths at locations between the stations. This has been challenging because an ultra-low gradient makes interactions between meteorology, vegetation, topology, and hydrology complex. Linear techniques such as interpolation and ordinary least-squares regression have under-performed because of the system's non-linear dynamics. This paper presents an alternative approach that employs artificial neural network (ANN) models to perform multivariate, non-linear interpolation between gauging stations. Using a combination of static and dynamic variables, predictions are generated in two modeling steps. The dynamic variables were 30-month time series of daily water depths at 16 stations and water levels (measured to the National Geodetic Vertical Datum of 1929) at 3 other stations. Static variable values were obtained from a previously developed GIS application having a 400-square-meter grid. Values included coordinates of cell centroids and percentages of vegetation types (slough, prairie, sawgrass, or upland) for approximately 2,300 cells, covering 370 square kilometers. The first ANN model interpolates mean water depths (for the period of record) from input static variables and mean water depths and levels at the gauging stations. The second ANN model predicts day-today variability about the interpolated means using a combination of static and dynamic variable inputs. A complete interpolation at a given cell is computed by summing the outputs of both models. Five of the water-depth gages were withheld from model development to validate model accuracy. Prediction accuracy was greatly improved, resulting in an average root-mean square prediction error at validation stations of only 3 centimeters (0.1 foot), or 4 percent of the dynamic range. Figure 1. Map showing study area, gauging stations and the EDEN grid, proposed index stations, final model stations and validation stations.