Hourly Water Level Prediction using Artificial Neural networks in Wetlands of Baldwin County, AL

Coastal wetlands are among the most important ecosystems in terms of services they provide (e.g., water quality improvement, water storage, habitat), but they are also among the most vulnerable. The most pressing issue leading to the loss of functional wetlands is increasing urban growth along the coast. In Alabama, headwater-slope wetlands are predominantly forested wetlands at the headwaters of creeks in the coastal plain. They tend to occur on relatively flat terrain where gradual slopes move water slowly through them as shallow groundwater discharge. They are potentially critical components to the landscape because they occur at the interface of uplands and coastal creeks. As a result, these wetlands intercept draining waters and improve water quality before reaching local bay and oceanic waters. To assess the impact of land use/cover on the functioning of headwater-slope wetlands and their associated functions, ground water levels in 15 headwater-slope wetlands in Baldwin County, AL were monitored over a two year period. However, due to various reasons each site has periods of missing data, varying from one to several months. In order to have a better understanding of the hydrology in these wetlands and relate them to the land use/cover conditions of their watersheds, we had to fill the data gaps. Thus, we developed an artificial neural network (ANN) based model to predict the groundwater levels in these wetlands. We picked 3 wetland sites to test the developed model. Spearman's rank correlation was used to find appropriate input vectors for training and testing phases of the ANN model. The most effective input for all the stations was water level data from the nearest stations and antecedent precipitation. A comparison of results from all stations declared that ANN trained with scaled conjugate gradient (scg) algorithm encompassed by two nearest stations and antecedent precipitation values with two (hourly) lags outperformed other training algorithms and input combinations. For the three sites, R2 and RMSE values at the validation phase ranged between 0.85 to 0.98 and 4.3 to 3.0 cm, respectively. Results also showed that finding the optimal training and testing datasets has a considerable effect on developing a satisfactory model for data mining problem at an hourly time scale.

Keywords: Ground water level, Wetlands, ANNs, Spearman's rank correlation