A Deep Learning-Based Ensemble Surface Energy Balance Modeling Approach
to Monitor Crop Water Use and Water Stress in drylands
Abstract
Most remote sensing-based surface energy balance (SEB) models are
limited by data availability and physical constraints to fully capture
the non-linear and temporally varying nature of atmospheric,
biophysical, and environmental controls on evapotranspiration (ET). As
such, currently, no single SEB model is considered to work best under
all conditions particularly in irrigated croplands where surface
moisture conditions could change dramatically in a short amount of time.
Hence, irrigation water management based on a single remotely sensed ET
model is often required to cope with model limitations and data latency
issues, which could lead to unsustainable and unreliable accounting of
water use over time. The recent inception of ensemble-based ET modeling
takes the advantage of the strengths of the several SEB models under
different conditions and is found to perform better as compared to an
Individual model. Yet, challenges remain in how high-temporal ET outputs
from different models are accurately assembled in a way that yields the
most reliable estimates of ET across any environmental and surface
conditions. Specifically, existing simple or Bayesian average and
machine learning-based ensemble approaches have not been able to
optimally utilize the comprehensive suite of existing SEB models and the
availability of multiple remotely sensed datasets. Here, we discuss the
utility of convolutional neural networks (CNNs) to assemble the outputs
from a host of SEB models that can robustly capture the non-linear
dynamics of ET under all conditions. We will also discuss the advantage
and potential limitations of using the CNN-based ensemble ET modeling
framework with respect to the individual, simple or Bayesian average,
and other machine learning approaches and their implications for use in
allocating water use across critically dry regions. Several ensemble
models will be trained using eddy covariance flux data globally and will
be evaluated based on their ability to estimate ET from MODIS and
Landsat sensors with both individual and fused products and minimal
weather inputs. The results can provide useful insights into how
multiple datasets and SEB models could be optimally utilized to
accurately monitor crop water status and support sustainable water
resource management in drylands.