3.3 Comparison of 1 (pond-only) versus 2 (paired pond-control)
logger design
Under the wet state criteria defined above, 3-state HMMs for the
pond-only model accurately predicted inundation states for 92% of site
visits for the 14 ponds. The paired pond-control model, which used
combined data from the loggers inside and outside of each pond,
accurately predicted inundation states for 90% of sites visits. Most of
the incorrect state predictions were likely due to sediment or
additional debris accumulating within the rugged housing units, which
was more likely to affect precision of drying dates rather than initial
inundation timing.
Models using temperature data from pond loggers alone predict inundation
timing that closely aligned with those using paired pond-control logger
data, indicating that a single logger design may be sufficient to
capture inundation timing of longer-duration events (Table S5, Figure
4). However, control logger data may help alleviate some of the wet
state false-positives, particularly when the standard deviation of daily
air temperature is relatively low or issues such as sediment in rugged
housing units occur. For example, earlier inundation dates are predicted
for several ponds by the pond-only model relative to the paired
pond-control model. This may be a true wet state, or the coincident low
temperature standard deviations measured by the control loggers may have
simply resulted in lower variance in the temperature on those days. For
site T13, the state was correctly predicted as wet by the paired
pond-control model, but not by the pond-only model in April 2019. While
we did observe water in the pond at this time, the water level was just
at the base of the rock pile covering the logger housing, which may
explain the discrepancies between the models. In cases such as this when
shallow water is present, the 2-logger design may help to increase the
probability of detecting inundation. Predicting pond drying may require
an array of pond loggers situated at different heights within the pond
to capture this fine-scale variation or a different type of sensor, such
as pressure transducers. But considerations exist for the pond-control
logger model as well. For example, pond-only models predicted wet states
for most ponds in the winter months (between December and February) that
were not predicted by the paired pond-control models. The relatively low
tSDs of the loggers in the winter months may be due to snow accumulation
on top of the control loggers.