Machine Learning Forecasts of Hourly and Daily Watertable Levels in a
Wet Prairie.
Abstract
Wet Prairies are unique communities occurring in isolated, saturated
depressions that typically have standing water from late fall to late
spring and then dry up and are burned by lightning or human-set fires in
summer. Wet prairies are highly sensitive to precipitation and
evapotranspiration patterns as they are highly dependent on shallow
groundwater levels. Three years (May 2015-May 2018) of hourly watertable
data recorded by a datalogger in a piezometer, and hourly precipitation,
temperature, relative humidity, and wind speed data from the Toledo
Express Airport (Ohio) weather station were used as inputs for a
nonlinear autoregressive neural network with exogenous inputs (NARX)
model. The data was used to train the model and to forecast watertable
levels at an hourly and a daily time step. The NARX model used a
Levenberg-Marquardt learning algorithm along with a combination of input
delay, feedback delay, and hidden layers. The training set encompassed
70% of the total dataset, with cross-validation and testing covering
the remaining 30% of the total dataset split equally among them. The
model was trained with the hourly data and then the hourly data was
aggregated to a daily record, and a new NARX model trained using this
new time-step. The intention was to see if a much smaller data record,
but typically all that is needed by a land manager, was capable of
producing a satisfactory watertable forecast. The NARX model’s
performance was assessed using R2, RMSE, and the Nash-Sutcliffe index.
The NARX model provided successful short-term forecasts (6 months) for
hourly and daily temporal resolutions. The R², RMSE, and Nash-Sutcliffe
for the hourly testing period were 0.85, 0.08, and 0.85 respectively.
For the daily testing period the R², RMSE, and Nash-Sutcliffe were 0.91,
0.07, and 90 respectively. The NARX model was not able to predict a
sudden increase in water table levels due to a large snowmelt event in
2018, but this is not surprising as the model was not trained using snow
melt events or a snow depth variable. Regardless of its current
limitations, land managers could use this NARX model to better
understand watertable patterns in wet prairies, one of the main drivers
of this natural community.