Soil Moisture Prediction Model from ERA5-Land Parameters using a
Deep Neural Networks
Daouda Diouf
1
, Carlos Mejia
2
and Djibril Seck
3
1
Laboratoire de Traitement de l’Information (LTI– ESP), Université Cheikh Anta Diop de Dakar, Senegal
2
IPSL/LOCEAN, Sorbonne Université, Paris, France
3
Université Cheikh Anta Diop de Dakar, Senegal
Keywords: Deep Neural Network, Soil Moisture, AdaGrad, ERA5-Land, CCI-ESA.
Abstract: In a global context of scarcity of water resources, accurate prediction of soil moisture is important for its
rational use and management. Soil moisture is included in the list of Essential Climate Variables. Because of
the complex soil structure, meteorological parameters and the diversity of vegetation cover, it is not easy to
establish a predictive relationship of soil moisture. In this paper, using the large amounts of data obtained in
West Africa, we set up a deep neural network to establish an estimation of soil moisture for the two first layers
and its prediction temporally and spatially. We construct deep neural network model which predicts soil
moisture layer 1 and layer 2 multiple days in the future. Results obtained for accuracy training and test are
greater than 93 %. The mean absolute errors are very low and vary between 0,01 to 0,03 m
3
/m
3
.
1 INTRODUCTION
The most important resource for the survival and
development of the earth's population is water
(Schlesinger, 2014). The soil moisture is the amount
of water level present in the top layers of the soil. The
soil moisture interacts and affects with atmosphere by
evaporation and transpiration (Kaleita et al., 2014;
Seneviratne et al., 2010). Temperature variability and
heatwaves have large dependence on soil moisture
feedback on evapotranspiration (Miralles et al., 2014;
Mueller and Seneviratne, 2012).
Many instruments and procedures can be used to
measure the soil moisture. Then, when the soil
moisture measurements are done by using
gravitimetric and volumetric procedures, it is called
direct method. Indirect method involves using
instrument like tensiometers, gypsum blocks, and
neutron probes.
The high correlation between soil moisture and
reflection spectrum of soil involve that many
researchers used remote sensing data to infer soil
moisture. The reflectance of soil in visible and
infrared bands is highly related to the soil colour,
texture, surface roughness and crusting, composition
and organic matter.
Reanalysis, that combines model data with
observations from across the world into a globally
complete and consistent dataset using the laws of
physics, offers spatial and temporal coverage
(Balsamo et al., 2015).
A good knowledge of soil moisture prediction can
be helpful in irrigation water management. It
involves better estimation of fertilizers and other
input, and better assessment of need and availability
of soil water level for crop cultivation. Thus, it is
necessary to be able to accurately predict soil
moisture in order to be able to save water, especially
for farmers.
Empirical formulas, linear regression, and neural
networks are currently the most widely used methods
for predicting soil moisture.
By the use of daily meteorological records, soil
physical properties, basic crop characteristics and
topographical data, Vahedberdi et al., (2009)
developed the Bridge Event And Continuous
Hydrological (BEACH) modelling to provide timely
information on the spatially distributed soil moisture
content over a given area without the need for
repeated field visits.
Using a soil moisture, precipitation and drought
prediction model, it was possible to predict drought
in a soil several days into the future (Chen et al,
2014).
Cai et al., (2019) use a deep learning regression
network, built with a two-layer hidden layer, to
Diouf, D., Mejia, C. and Seck, D.
Soil Moisture Prediction Model from ERA5-Land Parameters using a Deep Neural Networks.
DOI: 10.5220/0010106703890395
In Proceedings of the 12th International Joint Conference on Computational Intelligence (IJCCI 2020), pages 389-395
ISBN: 978-989-758-475-6
Copyright
c
2020 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
389
establish a predictive model between meteorological
parameters and soil moisture at a depth of 20 cm in
the Yanqing area (Beijing, China) with accuracy of
98%.
The objective of this work is to accurately predict
the soil moisture level multiple hours in advance by
using deep neural network regression. With few
parameters easy to measure and easy to access, the
challenge in this work is to successfully predict the
evolution in time and space of soil moisture.
The reason for choosing deep learning is that with
these methods it was possible to improve the accuracy
of soil prediction due to its non-linearity and structure
complexity (Veres et al., 2015; Cai et al., 2019).
2 DATASET
We use ERA5-Land hourly dataset with ~9km grid
spacing. ERA5-Land has been produced by replaying
the land component of the ECMWF ERA5 climate
reanalysis. Thus, ERA5-Land is forced by the
atmospheric analysis of ERA5 and hence
observations indirectly influence the simulations.
This dataset is taken in an area of the West Africa,
between 6°N and 24°N and -17°W and 34°W.
West Africa's climate is characterized by a strong
latitudinal rainfall gradient that determines
production systems. It is also characterized by
dramatic fluctuations in rainfall patterns on multi-
decadal time scales, amplifying the already
substantial annual rainfall variability. These include
sub-humid, semi-arid and arid zones.
The climatology of the average annual
precipitation cycle can be summarized in a few main
phases. The first rains appear on the coasts of the Gulf
of Guinea (5°N) in March; they then increase in
intensity during the months of April and May; during
the month of June, the zone of heavy rainfall moves
rapidly towards latitudes close to 10°N (Sultan and
Janicot, 2000), remaining almost stationary at this
position until the end of August, a period which
corresponds to the short dry season in the Guinean
zone. Rainfall decreases in August, linked to the
relative atmospheric stability on the coasts of the Gulf
of Guinea resulting from the drop-in ocean
temperatures and a divergence in specific humidity
(Philippon and Fontaine, 2002). Finally, there is a
gradual withdrawal of the rainy zone towards the
coasts between September and November, a period
that corresponds to the beginning of the second
passage of the ITCZ along the coasts (second rainy
season).
The learning dataset describes eight (08) variables
and two (02) moisture soil layer. These 08 features
are noted by x and the volumetric soil moisture.
Volumetric soil moisture is expressed in m
3
.m
3
.
Features x are composed of five meteorological
data such as 2 metre temperature (t2m), 2 metre
dewpoint temperature (d2m), total precipitation (tp),
10m u-component of wind (u10) and 10m v-
component of wind (v10); two parameters related to
soil properties such as evaporation from bare soil
(evabs) and surface sensible heat flux (sshf); and the
initial soil moisture (smli).
Soil moisture is localized in ERA5-Land in 4
layers with depths of 0.07 (0–0.07), 0.21 (0.07–0.28),
0.72 (0.28–1.00) and 1.89 (1.00–2.89) m. The first
two layers are of interest to us in this study.
For each ERA5-Land day, we take measurements
at 00 h and 12 h. These measurements concern the
years from 2012 to 2013 for the months from July to
November. This gives a matrix with a dimension of
10 x130000.
For validation dataset, we used combined various
single-sensor active and passive microwave soil
moisture from Climate Change Initiative (CCI) of the
European Space Agency (ESA). These level 3 (super-
collated: L3S) dataset are observations combined
from multiple instruments into a space-time grid. The
soil moisture data for the combined product are
provided in volumetric units [m3.m-3]. The products
come, among others, from sensor as Scanning
Multichannel Microwave Radiometer (SMMR)
onboard Nimbus-7, Tropical Rainfall Measuring
Mission (TRMM), the Advanced Scatterometer
(ASCAT) onboard the Meteorological Operational
satellite program (MetOp), the Special Sensor
Microwave Imager (SSM/I), the Advanced
Microwave Scanning Radiometer — Earth Observing
System (AMSR-E) on-board the Aqua satellite.
3 METHODS
The main objective of machine learning is to estimate
the unknown relationship between input and target
parameters using known examples. For numerical
targets, the tasks become a supervised learning. The
objective of supervised learning is to build
relationships and dependencies model between the
target prediction output and the input features such
that we can later predict the output values for new
data based on the model.
Suppose
N
n
nn
yx
1
,
to be the training dataset
with X being the input space and Y being the output
space. The objective at the moment is to seek a
NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications
390
function f: X
Y from a hypothesis space that
minimizes the loss associated. The best fit to the
underlying function can be chosen by minimizing a
cost function.
Consider
i
y
ˆ
the predicted value,
i
y
the true
value, and the average value, the performance of a
model can be measured by:
Mean Absolute Error (MAE):
n
ii
yy
n
1
)
ˆ
(
1
(1)
R Squared (R
2
):
i
i
i
ii
yy
yy
2
2
)(
)
ˆ
(
1
(2)
To build supervised learning model, several
algorithms, which are developed in different
mathematical backgrounds, exist. We can denote,
linear regression, ridge regression, decision trees, K-
Neighbors regression, Support vector regression,
neural networks (Diouf and Seck, 2019).
For this study, we are taken a neural network
method. A neural network is a mathematical model
used as nonlinear statistical tools in modeling
complex relationships between inputs and outputs.
We opt to a two-hidden layer regression neural
network. The output of the previous layer is the input
of the next layer. It is a deep neural network
regression and its mathematical structure is composed
by:
- An input layer for which the number of nodes is
equal to the number of input parameters.
- Hidden layers node composed of neurons.
- The regression model output layer. The output of the
previous hidden layer is multiplied by the weight and
is added to a bias on the output node to obtain the
regression prediction value.
We use a single model to predict soil moisture for
layer 1 and layer 2, so-called 2NNL2. This model is a
succession of two networks to form a unique model.
The first network has as input the eight parameters
and as output the soil moisture of layer 1. The second
network have as input the same inputs of the previous
network plus the output of network 1. The output is
the soil moisture of layer 2.
We use five models to predict the soils moisture
level multiple days in advance.
Model 1: The output data of network 1 (sml1) is
measured two (02) days after the input data. The
output data of network 2 (sml2) is measured three (03)
days after the input data and one (01) day after the
sml1.
Model 2: The output data of network 1 (sml1) is
measured three (03) days after the input data. The
output data of network 2 (sml2) is measured four (04)
days after the input data and one (01) day after the
sml1.
Model 3: The output data of network 1 (sml1) is
measured four (04) days after the input data. The
output data of network 2 (sml2) is measured five (05)
days after the input data and one (01) day after the
sml1.
Model 4: The output data of network 1 (sml1) is
measured five (05) days after the input data. The
output data of network 2 (sml2) is measured six (06)
days after the input data and one (01) day after the
sml1.
Model 5: The output data of network 1 (sml1) is
measured six (06) days after the input data. The
output data of network 2 (sml2) is measured seven
(07) days after the input data and one (01) day after
the sml1.
This means that for each model, the inputs of network
2 are the same inputs of network 1 plus output of
network 1 (sml1).
After many attempts, all these models’ structure
was determined to be 8-150-80-1 followed by 8-100-
50-1 respectively for network 1 and network 2.
We train and optimize Model 1, Model 2, Model
3, Model 4 and Model 5.
Several algorithms can be used for optimization.
Here we choose Adaptive Gradient Algorithm
(AdaGrad) as an optimization algorithm (Duchi et al.,
2011). AdaGrad is an optimization algorithm for
gradient-based optimization. AdaGrad performs
gradient descent with a variable learning rate.
Parameters associated with infrequent features are
adapted with large gradients and parameters
associated with frequently occurring features perform
small gradients. Adagrad thus improves on SGD, or
stochastic gradient descent, with a per-node learning
rate scheduler built into the algorithm.
To optimize gradient descent at time-step t,
t
g
,
an objective function
J
is minimized by updating
a parameter
. The equation of the parameter is:
t
t
tt
g
GdiagI
)(
1
(3)
where
t
is the parameter to be updated at time-step
t, η is the learning rate, ε is some small quantity that
used to avoid the division of zero, I is the identity
Soil Moisture Prediction Model from ERA5-Land Parameters using a Deep Neural Networks
391
matrix,
)(
t
Gdiag
is a diagonal matrix containing the
squares of all previous gradients,
t
g
is the vector of
gradients for the current time-step and can be
expressed, for each training example
i
x
and label
i
y
, by:
n
t
ii
t
yxJ
n
g
1
),,(
1
(4)
The accuracy on the learning set is 93.8% and the
validation accuracy is 92.5% for all models. The
mean absolute error turn around 0.015 m
3
/m
3
for
training phase and 0.02 m
3
/m
3
for validation phase.
Table 1 summarizes performances measures for all
models. We notice that the soil moisture retrieved
from training features and its real values are quite
good for layer 1 and layer 2. This figure gives us an
idea of the accuracy of the model in reproducing the
training dataset.
Figure 1: A two connected two-hidden layer regression neural network (2NNL2).
Table 1: Performance measures of 2NNL2 models.
Train mae Test mae Train loss Test loss Train R
2
Test R
2
Model 1
Output 1 0.010 0.013 0.0002 0.0005 98.37% 98.43%
Output 2 0.023 0.026 0.0010 0.0013 93.8% 93.7%
Model 2
Output 1 0.013 0.015 0.0004 0.0006 97.7% 97.2%
Output 2 0.023 0.026 0.0011 0.0014 93.8% 92.7%
Model 3
Output 1 0.015 0.018 0.0005 0.0007 97.1% 97.1%
Output 2 0.023 0.027 0.0010 0.0015 93.8% 92.5%
Model 4
Output 1 0.017 0.020 0.0006 0.0009 96.8% 96.5%
Output 2 0.024 0.027 0.0011 0.0013 93.9% 92.7%
Model 5
Output 1 0.018 0.021 0.0006 0.0010 96.4% 96.4%
Output 2 0.024 0.027 0.0011 0.0014 93.6% 92.5%
NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications
392
Figure 2: Scatter plot (left) and relative error (right) between predicted and measured values for sml1 in 12 July, 2012.
Figure 3: Scatter plot (left) and relative error (right) between predicted and measured values for sml2 in 13 July, 2012.
4 RESULTS
In the training phase, the dataset selected concern
sample less than 20% of all measured values from
July to December 2012. We compared two data sets
of soil moisture that did not participate in training
phase to the measure from ERA5-Land at the same
date.
Using July 10, 2012 input parameters, we predict
sml1 and sml2 two days and three days in the future,
respectively, i.e. on dates of 12 and 13 July, 2012.
Figure 2 and figure 3 show comparisons of soil
moisture layer predicted and measured. We can
notice that the prediction model was able to retrieve
the soil moisture very faithfully. The accuracies of
scatter diagrams are 95.6% and 94.4% respectively
for sml1 and sml2.
The global mean absolute error between the two
data sets is quite small: 0.03 m
3
/m
3
. Then, the sml1
retrieval from the Era5-Land features by using neural
network are obtained with good accuracy. In the
construction of the model, the output sml1 of the first
stage is part of the input of the second stage which
models the sml2. This means that a good estimate of
the output of stage 1 will lead to a good estimate of
stage 2. The contrary will also cause the opposite
effect. These comparisons on dataset that not
participate to the training phase between observed
and estimated show the generalization capability of
the built model.
Soil moisture obtained from Climate Change
Initiative of the European Space Agency (CCI-ESA),
which are combination of measurements from various
single-sensor active and passive microwave, is used
to validate mainly our model and occasionally the
ERA5-Land data.
Comparison between the sml1 predicted two days
in the future from model with using the ERA5-Land
parameters reanalysis (a) and the measured ESA-CCI
sml1 (b) on July 10, 2012 can be seen in figure 4. The
soil moisture prediction two days in the future was
compared with measurements from ESA-CCI data. A
correlation of 87% and a mean absolute error of 0.05
m
3
/m
3
were obtained. For the prediction made two
Soil Moisture Prediction Model from ERA5-Land Parameters using a Deep Neural Networks
393
days in the future for layer 1 in figure 5 (date of July
12, 2012), we also note good results with a correlation
coefficient of 85% and a mean absolute error of 0.06
m
3
/m
3
. For the two figures shown below, we note that
the trends are the same in part (a) and part (b).
However, the intensities of soil moisture predicted
from Era5-Land features are on average 0.05 m
3
.m
-3
higher than those measured from CCI-ESA.
Figure 4: Map of the soil moisture layer 1 predicted from ERA5-Land features with 2NNL2 (a) and CCI-ESA observations
data (b) in 10 July, 2012.
Figure 5: Map of the soil moisture layer 1 predicted from ERA5-Land features with 2NNL2 (a) and CCI-ESA observations
data (b) in 12 July, 2012.
NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications
394
5 CONCLUSIONS
In this study, dataset from ERA5-Land were used to
build a prediction model by using a deep neural
network able to evaluate further soil moisture in the
first two layers. The built model, so-called 2NNL2,
which is a succession of two-hidden layers, retrieved
successfully soil moisture layer 1 and layer 2 for two
to seven days in the future. We have analyzed the
performance of the model by comparing soil moisture
estimated from ERA5-Land features to CCI-ESA soil
moisture. We denoted that results are satisfying with
low mean absolute error and high correlation.
REFERENCES
Balsamo, G., Albergel, C., Beljaars, A., Boussetta, S., Brun,
E., Cloke, H., Dee, D., Dutra, E., Muñoz-Sabater, J.,
Pappenberger, F., de Rosnay, P., Stockdale, T., Vitart,
F., 2015. ERA-Interim/Land: a global land surface
reanalysis data set. Hydrol. Earth Syst. Sci. 19, 389–
407.
Cai Y., Zheng W., Zhang X., Zhangzhong L., Xue X.,
(2019) “Research on soil moisture prediction model
based on deep learning”. PLoS ONE 14(4): e0214508.
https://doi.org/10.1371/journal.pone.0214508
Chen XF, Wang ZM, Wang ZL, Li R. Drought evaluation
and forecast model based on soil moisture simulation.
China Rural Water and Hydropower, 2014(05): 165–
169.
Diouf D. and Seck D., (2019) “Modeling the chlorophyll-A
from sea surface reflectance in West Africa by deep
learning methods: A comparison of multiple
algorithms”. International Journal of Artificial
Intelligence & Applications; 2019; Vol 10(6); Pages
33-40
Kaleita A. L., Tian L. and Yao H., “Soil moisture estimation
from remotely sensed data,” in American Society of
Agricultural Engineers Annual International meeting,
Las Vegas, NV, July 27-30, 2003.
Duchi, J., Hazan, E., & Singer, Y. (2011). Adaptive
Subgradient Methods for Online Learning and
Stochastic Optimization. Journal of Machine Learning
Research, 12, 2121–2159.
LeCun Y. A., Bottou L., Orr G. B., and Müller K.-R.,
“Efficient backprop,” in Neural Networks: Tricks of the
Trade, pp. 9–48, Springer, Berlin, Germany, 2012.
Miralles, D.G., Teuling, A.J., van Heerwaarden, C.C., Vila-
Guerau de Arellano, J., 2014. Mega-heatwave
temperatures due to combined soil desiccation and
atmospheric heat accumulation. Nat. Geosci. 7, 345–
349.
Mueller, B., Seneviratne, S.I., 2012. Hot days induced by
precipitation deficits at the global scale. Proc. Natl.
Acad. Sci. 109 (31), 12398–12403.
Philippon N. et Fontaine B., 2002: The relationship
between the Sahelian and previous 2nd Guinean rainy
seasons: a monsoon regulation by soil wetness?
Annales Geophysicae, 20(4), 575-582.
Schlesinger W H, Jasechko S. Transpiration in the global
water cycle. Agricultural and Forest Meteorology,
2014, 189: 115–117.
Seneviratne, S.I., Corti, T., Davin, E.L., Hirschi, M., Jaeger,
E.B., Lehner, I., Orlowsky, B., Teuling, A.J., 2010.
Investigating soil moisture-climate interactions in a
changing climate: a review. Earth Sci. Rev. 99, 125–
161.
Sultan B. et Janicot S., 2000: Abrupt shift of the ITCZ over
West Africa and intra- seasonal variability. Geoph. Res.
Lett., 27, 3353-3356.
Vahedberdi S., Saskia V. and Stroosnijder L. “A simple
model to predict soil moisture: Bridging Event and
Continuous Hydrological (BEACH) modelling”,
Environmental Modelling & Software Volume 24,
Issue 4, April 2009, Pages 542-556
Veres M, Lacey G, Taylor G W. Deep learning
architectures for soil property prediction. Computer and
Robot Vision (CRV), 2015 12th Conference on. IEEE,
2015: 8–15.
Soil Moisture Prediction Model from ERA5-Land Parameters using a Deep Neural Networks
395