Using deep learning for multivariate mapping of soil with quantified
uncertainty
Wadoux, A. M. J. C.
This is a "Post-Print" accepted manuscript, which has been published in "Geoderma"
This version is distributed under a non-commercial no derivatives Creative Commons
) user license, which permits use, distribution, and
reproduction in any medium, provided the original work is properly cited and not
used for commercial purposes. Further, the restriction applies that if you remix,
transform, or build upon the material, you may not distribute the modified material.
Please cite this publication as follows:
Wadoux, A. M. J. C. (2019). Using deep learning for multivariate mapping of soil with
quantified uncertainty. Geoderma, 351, 59–70.
https://doi.org/10.1016/j.geoderma.2019.05.012
A
ccepted
Using deep learning for multivariate mapping of soil with
quantified uncertainty
Alexandre M.J.-C. Wadoux
a,
∗
a
Soil Geography and Landscape group, Wageningen University & Research
Abstract
Digital soil mapping (DSM) techniques are widely employed to generate soil
maps. Soil properties are typically predicted individually, while ignoring the in-
terrelation between them. Models for predicting multiple properties exist, but
they are computationally demanding and often fail to provide accurate descrip-
tion of the associated uncertainty. In this paper a convolutional neural network
(CNN) model is described to predict several soil properties with quantified un-
certainty. CNN has the advantage that it incorporates spatial contextual infor-
mation of environmental covariates surrounding an observation. A single CNN
model can be trained to predict multiple soil properties simultaneously. I further
propose a two-step approach to estimate the uncertainty of the prediction for
mapping using a neural network model. The methodology is tested mapping six
soil properties on the French metropolitan territory using measurements from
the LUCAS dataset and a large set of environmental covariates portraying the
factors of soil formation. Results indicate that the multivariate CNN model
produces accurate maps as shown by the coefficient of determination and con-
cordance correlation coefficient, compared to a conventional machine learning
technique. For this country extent mapping, the maps predicted by CNN have
a detailed pattern with significant spatial variation. Evaluation of the uncer-
tainty maps using the median of the standardized squared prediction error and
accuracy plots suggests that the uncertainty was accurately quantified, albeit
slightly underestimated. The tests conducted using different window size of in-
put covariates to predict the soil properties indicate that CNN benefits from
using local contextual information in a radius of 4.5 km. I conclude that CNN
is an effective model to predict several soil properties and that the associated
uncertainty can be accurately quantified with the proposed approach.
Keywords:
Pedometrics, Prediction intervals, Uncertainty quantification,
Convolutional neural network, Machine learning, LUCAS
∗
Corresponding author : Soil Geography and Landscape group, Wageningen University,
Droevendaalsesteeg 3, 6708 BP Wageningen, the Netherlands
Email address:
alexandre.wadoux@wur.nl
(Alexandre M.J.-C. Wadoux)
Preprint submitted to Geoderma
May 23, 2019
A
Highlights
• A convolutional neural network is used for soil mapping.
• Soil texture, organic carbon, pH and nitrogen are predicted using a single
model.
• The model is trained on topsoil LUCAS data over France.
5
• A two-step method to quantify the uncertainty of the prediction is tested.
• The method for uncertainty quantification is applicable to any neural net-
work model.
1. Introduction
Many agronomic and environmental activities require accurate information about
10
the spatial distribution of soil properties. This information is commonly gen-
erated by digital soil mapping (DSM) techniques, whose framework has been
formalized by the publication of McBratney et al. (2003). In DSM, prediction is
made by exploiting the empirical quantitative relationship between a measured
soil properties and one of several environmental covariates chosen to portray the
15
factors of soil formation. The factors correspond to
s
: soil,
c
: climate,
o
: organ-
ism/vegetation,
r
: relief/topography,
p
: parent material,
a
: age and
n
: spatial
position, which motivate the
scorpan
spatial model of soil variation. Usually,
a model is built for predicting each property individually. This can lead to in-
consistent prediction (Heuvelink et al., 2016) and rapid increase of computing
20
intensity as the number of models (and therefore parameters to estimate) grows.
For example, predicting soil clay, silt and sand would surely benefit from using a
common model so as to avoid producing a unrealistic map of soil texture. This
has been recognized in many previous DSM studies (e.g. Akpa et al., 2014).
Several methods exist for simultaneous prediction of soil properties, such as co-
25
kriging (Goovaerts, 1997), regression co-kriging (Heuvelink et al., 2016) and
structural equations modelling (SEM) (Angelini et al., 2017). These linear
methods model the interrelation between properties explicitly but are compu-
tationally demanding when the size of the observation dataset is large or the
number of properties to predict increases. In addition, they rely heavily on
30
rigid statistical assumptions about the distribution of the soil properties. As
an alternative, non-linear techniques such as machine learning have garnered
wide interest during the past decade. Hengl et al. (2018) have promoted ran-
dom forest for multivariate prediction of soil properties, while Xu et al. (2013)
have adapted support vector machine for predicting many dependent variables
35
2
A
ccepted
simultaneously. Despite those examples, the use of machine learning techniques
for multivariate soil mapping have been largely unexplored.
Recently, deep learning (DL) models have shown great potential for soil map-
ping. Deep neural network as employed by Behrens et al. (2018) produced
more accurate predictions compared to random forest. Wadoux et al. (2019)
40
and Padarian et al. (2019) used a convolution neural network (CNN) model for
mapping soil organic carbon. The authors showed how a single CNN model can
be used for predicting at multiple depths, and how the prediction accuracy sig-
nificantly increased when compared to maps produced by either random forest
or cubist regression trees. CNN has the advantage that it uses the contextual
45
information contained in the vicinity of a location by taking a local representa-
tion of the input covariates. Mapping using the spatial domain of the covariates
is not new and several approaches have been developed, such as spatial filters
using wavelet (Lark et al., 2004) or multiscale analysis (Miller et al., 2015).
However, while these approaches contextualize the spatial information supplied
50
to the model, they rely on either subjective modeller’s decision or heavy covari-
ates pre-processing, which hamper their use for predicting soil properties in an
operational context.
Prediction is not the only interest of map users. Quantifying prediction uncer-
tainty is as important as the prediction itself (Wadoux et al., 2018). Padarian
55
et al. (2019) derived confidence intervals by training several CNN models on
bootstrap samples of the input data. A confidence interval reflects the un-
certainty around the mean prediction values. In soil mapping, we are rather
interested in prediction intervals (Heuvelink, 2014), i.e. the range that is likely
to contain the value yet to be observed. As consequence, a prediction interval is
60
always wider than a confidence interval. Estimating only the latter certainly un-
derestimates the total uncertainty of the prediction. For neural network models,
several solutions to obtain prediction intervals have been proposed, such as the
Delta, Bayesian or bootstrap plus variance estimate methods (Khosravi et al.,
2011). To the best of my knowledge, quantifying prediction uncertainty of a
65
neural network model has been yet disregarded in DSM studies.
The objectives of this study were to use a single CNN model for multivariate soil
mapping and to quantify the uncertainty of the predictions. The methodology
is tested in a potential application scenario, mapping topsoil clay, silt, sand,
organic carbon, total nitrogen and pH in CaCI
2
solution over France. The
70
predicted soil maps were validated and the uncertainty was estimated of each
soil property.
2. Methodology
2.1. Convolutional neural network
Topsoil properties of interest
z
s
i
at location
s
i
(
i
= 1
, . . . , n
;
s
i
∈ F
)
for
p
soil
75
properties
z
p
(
p
= 1
, . . . , P
)
in the study area
F
are modelled by a convolutional
3
A
ccepted
neural network (CNN):
z
s
i
=
f
(
X
s
i
;
θ
) +
ε
s
i
(1)
where
X
is a 3-D input matrix of size
c
×
w
×
h
which contains
c
environmental
covariates of dimension
w
×
h
centred at the measured soil property location
s
i
. The vector
θ
is a set of model parameters used by regression model
f
to
80
map non-linearly
X
→
z
and leaves room for a zero mean random error vector
ε
. Measurements of the soil properties are assumed independent and identically
distributed.
A CNN model is composed of several layers (Goodfellow et al., 2016), among
which an input layer supplying the images
X
to the network. The input layer is
85
connected to a hidden layer which in turn is connected to another hidden layer
or to an output layer. Each layer contains neurons which are independent within
the layer but connected to each neurons from the previous and to the next layer.
Hidden layers can be classified in three main categories, called convolutional,
pooling and fully connected layers.
90
A convolutional layer has the particularity that it performs a convolution be-
tween an input image and a filter. Convolutional layers are placed at the begin-
ning of the network, and take an image of a given size and number of channels
(i.e. number of covariates) as input and returns another image with a possibly
different size but same number of channels. For a given input image
X
and a
95
non-linear function (usually a rectified linear unit (ReLU),
φ
(
x
) = max(0
, x
)
),
the convolution outputs the image
X
0
by:
X
0
=
p
(
φ
(
W
∗
X
+
b
))
,
(2)
where
W
is a matrix of weights of size
J
k
×
J
k
−
1
, i.e. the number of neurons
J
in the current
k
layer times the number of neurons in layer
k
−
1
. The vector
b
are the neuron bias, “
∗
” is a convolutional operator over dimensions
w
and
h
of
100
X
and
p
(
.
)
is a pooling function which selects the maximum value in the input
image using a given window size (max-pooling). Note that convolutional and
pooling layers share weights across the
w
and
h
dimensions which greatly reduce
the number of parameters to be estimated. The last convolutional or pooling
layer returns an image
X
, which can be converted to a vector
x
(the flatten
105
operation) and provided as input to a fully connected layer which outputs
x
0
as
follows:
x
0
=
φ
(
Wx
+
b
)
,
(3)
where
φ
is the ReLU activation function or the linear activation
φ
0
(
x
) =
x
for
the output layer. From Eq. 2 and Eq. 3 it follows that the model parameters
θ
=
(
W
1
b
1
, . . . ,
W
L
b
L
)
for
k
= 1
, . . . , L
hidden layers. For notational convenience
110
from here on I drop the subscript
p
so that
z
p,s
i
=
z
s
i
.
Parameters
θ
can be estimated by training the CNN model to the dataset
D
=
{
(
X
s
i
,
z
s
i
)
. . .
(
X
s
n
,
z
s
n
)
}
by minimizing the mean squared error (MSE)
4
A
ccepted
as objective function, defined by:
MSE =
1
n
n
X
i
=1
(
z
s
i
−
ˆ
f
(
X
s
i
; ˆ
θ
))
2
.
(4)
Note that one objective function is computed for each soil property
p
, but only
115
one single multi-task CNN is trained. In this case, the objective function is
simply the average of all soil properties objective function. The Adam opti-
mizer (Kingma & Ba, 2015) is used to minimize Eq. 4. Adam computes the
derivative of the objective function with respect to the model weights and bias
to update their value in a process called backpropagation (LeCun et al., 1989).
120
The optimization process runs for a number of epochs. An epoch describes the
number of times the network sees the entire input dataset. During each epoch,
the entire dataset is shown to the network in small subsets shuffled at random,
called batches. The number of epochs as well as the batch size is chosen by the
user. Another hyperparameter is the learning rate of the optimizer, i.e. how
125
fast the optimizer moves the weights in the opposite direction of the gradient
after each update. A too small learning rate increases the computation time
to find the optimum of the objective function because the steps are small. If
the learning rate is too large training may not converge because the weights
oscillate.
130
2.2. Uncertainty quantification
To estimate the variance of the prediction, I use a two-step method called boot-
strap plus variance estimate (Khosravi et al., 2011). Recall from Eq. 1 that
errors are assumed to be statistically independent and identically distributed.
Using a predicted mean of the soil property of interest by regression model
135
ˆ
f
(
X
,
ˆ
θ
)
, shortly denoted
ˆ
f
hereafter, one can rewrite Eq. 1 as:
z
s
i
−
ˆ
f
s
i
= (
f
s
i
−
ˆ
f
s
i
) +
ε
s
i
.
(5)
The first term in the right-hand side of Eq. 5 is the difference between the
estimated model
ˆ
f
and true regression model
f
, which relates to the probability
distribution
p
(
f
|
ˆ
f
)
(called confidence interval). The left-hand side of Eq. 5 is the
difference between the measured values
z
p
and predicted values
ˆ
f
, which relates
140
to probability distribution
p
(
z
p
|
ˆ
f
)
(called prediction interval). Therefore, the
total variance of the prediction is formulated as (Khosravi et al., 2011):
σ
2
s
i
=
σ
2
ˆ
f
si
+
σ
2
ˆ
ε
si
,
(6)
where
σ
2
ˆ
f
is the error term due to model error, i.e. model miss-specification
and model parameter uncertainty and
σ
2
ˆ
ε
is the data error, i.e. the data noise
variance. Estimation of these two terms separately is presented in more detail
145
in the two next paragraphs.
5
A
1. The model error variance term is estimated by the bootstrap method
(Efron & Tibshirani, 1994), which builds an ensemble of CNN models,
each with a random initialization of the parameters
θ
. From the original
dataset,
B
training sets are randomly sampled with replacement, forming
150
D
B
b
=1
training datasets. Next,
B
CNN models are trained on bootstrapped
data
D
b
. The mean of
B
model outputs for a soil property
p
is computed
by:
ˆ
z
s
i
=
1
B
B
X
b
=1
ˆ
z
b,s
i
,
(7)
where
ˆ
z
b
is the prediction of the
b
th bootstrap model for the soil property
p
. The model error variance term can be estimated using the prediction
155
from
b
bootstrap models by:
σ
2
ˆ
f
si
=
1
B
−
1
B
X
b
=1
(ˆ
z
b,s
i
−
ˆ
z
s
i
)
2
.
(8)
Note that models are trained based on the minimization of objective func-
tion defined in Eq. 4.
2. The data error variance term is estimated following Nix & Weigend (1994)
by assuming normally distributed errors around
ˆ
f
. In this case, the CNN
160
model outputs two values in the final layer, corresponding to the predicted
mean
ˆ
z
and variance
σ
2
ε
of a Gaussian distribution. The least square re-
gression can be interpreted as maximum likelihood, by minimizing the neg-
ative log-likelihood criterion instead of Eq. 4, given by (Lakshminarayanan
165
−
log(
z
s
i
|
X
s
i
) =
log
σ
2
ε
si
2
+
(
z
s
i
−
ˆ
z
s
i
)
2
2
σ
2
ε
si
+ constant
.
(9)
The variance term in Eq. 9 is passed through the function
log(1 + exp(
.
)
to enforce the positivity constrain. A small variance term (10
−
6
) is also
added to ensure numerical stability (Lakshminarayanan et al., 2017). The
optimizer used is the same as for minimizing Eq. 4. Once the CNN model
trained, it is used to predict the variance term
σ
2
ε
at any location.
170
Assuming both terms independent and Gaussian enables to compute the total
variance as the sum of the two terms. The method has the main disadvantage
that it requires to build two separate models (one for estimating
σ
2
ε
and another
for
σ
2
ˆ
f
) and to train
B
+ 1
CNN models. Khosravi et al. (2011) note that
this method relies on the
B
bootstrap model estimates. Some of the bootstrap
175
samples may lead to a biased prediction. In consequence, the variance can be
underestimated resulting in narrow coverage probability.
6
A
ccepted
2.3. Quality of prediction and estimated uncertainty
For the vector
z
p
at
N
−
n
test locations where
n
is the number of calibration
location and
N
is the total number of sampling locations, the quality of the
180
prediction is quantified by the mean prediction error (ME), root mean squared
error (RMSE), amount of variance explained by the model (
r
2
) and concordance
correlation coefficient (CCC). The latter is derived as follows (Lawrence & Lin,
CCC =
2
ρ
0
σ
z
p
σ
ˆ
z
p
σ
2
z
p
+
σ
2
ˆ
z
p
+ (
µ
z
p
−
µ
ˆ
z
p
)
2
,
(10)
where
µ
p
and
σ
2
p
are mean and variance for the vector of true measurements
z
p
185
of soil property
p
or for the vector of predicted values
ˆ
z
p
. The value
ρ
0
represents
the correlation between
µ
z
p
and
µ
ˆ
z
p
. The CCC quantifies the agreement of the
predictions to the 1:1 line. Its optimal value is 1 and it can be negative.
The quality of the prediction error variance is quantified by the standardized
squared prediction error
δ
p
190
δ
p,s
i
=
(
z
s
i
−
ˆ
z
s
i
)
2
σ
2
ˆ
z
si
,
(11)
which should be distributed as
χ
2
with 1 degree of freedom so that its mean
δ
should be close to 1. Lark (2000) showed that its median
˜
δ
is more useful
because it is less sensitive to small or large values,
˜
δ
should be close to 0.455.
In addition, the prediction intervals coverage probability (PICP) is computed.
The PICP is the percentage of observations covered by a defined prediction
195
interval (Shrestha & Solomatine, 2008), in this case corresponding to a 90%
probability of occurrence. The PICP is calculated as follows:
PICP =
1
N
−
n
N
−
n
X
i
=
n
+1
P
i
·
100
,
(12)
for
P
i
=
(
1
if
q
(0
.
05)
≤
z
s
i
≤
q
(0
.
95)
,
0
otherwise,
(13)
where
q
(0
.
05)
and
q
(0
.
95)
are the lower and upper boundaries of 90% probability
of occurrence at location
s
i
.
200
Finally, a visual assessment of the quality of the estimated uncertainty is pro-
vided by an accuracy plot as first proposed by Deutsch (1997). Having normally
distributed error at each test location with known mean and variance allows
computing a symmetric interval around the predicted values by calculation of
the
(1
−
q
)
/
2
and
(1 +
q
)
/
2
quantiles, for a number of
q
intervals. One can count
205
the proportion of observed values at test location included for each
q
interval.
If the uncertainty is correctly modelled, the proportion of observations covered
7
A
ccepted
by a
q
interval at test locations is approximately equal to the value of
q
, for all
q
. The values of
q
can be plotted in a scattergram against the actual propor-
tion of observation in each
q
. Ideally all points in the plots are on the 1:1 line.
210
Deviation from the 1:1 line is due to overestimation or underestimation of the
modelled uncertainty, depending whether the points lie above or below the line,
respectively.
3. Case study
3.1. Study area and data
215
The methodology is tested on the metropolitan territory of France which is
about 543,965 km
2
excluding Corsica and other islands. France has a very di-
verse landscape and climate. The altitude ranges from 0 to more than 4,500 m
in the Alps. The climate is Mediterranean in the South and temperate in the
North, influenced by a West-East gradient of decreasing precipitation and tem-
220
perature due to the increasing distance from the Atlantic ocean. France is mostly
covered by soils from calcareous rocks such as the Rendzic, calcaric Leptosol and
Calcisols on the Champagne region or in Argonne. Soils from clayey sediments
such a Haplic and Vertisols are found in patches north of Massif Central and
in the southern Alps mountains with fertile loess soils (Luvisols) in the North.
225
Sandy soils such as Podzols are found in the Landes or Sologne while large areas
in the Massif Central and Brittany are covered by dystric Cambisols originating
from moderate weathering of different types of parent materials (Jones et al.,
2005).
In this study I used the soil measurements from the land use and cover area frame
230
statistical survey (LUCAS) covering the French territory. The LUCAS dataset
is a harmonized dataset of about 20,000 topsoil (0-10 cm) samples covering the
whole Europe. The LUCAS dataset has been used in many previous studies on
soil spatial distribution (e.g. Ballabio et al. (2016)). In this study are predicted
six soil properties from the LUCAS dataset covering France. The properties are
235
the soil organic carbon in g kg
−
1
, particle size fraction (clay, silt and sand in %),
pH in CaCI
2
solution and total nitrogen content in g kg
−
1
, denoted OC, clay,
silt, sand, pH and N hereafter. For more information about the LUCAS dataset
I refer to Tóth et al. (2013). In addition, a set of thirty-seven readily available
environmental covariates were assembled to represent the
scorpan
factors of soil
240
formation. Table 1 lists the covariates with their unit and original reference.
Any covariate which did not conform with the target grid resolution of 1 km
×
1 km was either resampled using bilinear interpolation or aggregated. Cat-
egorical covariates such as landforms and geology were transformed to dummy
variables to allow subsequent analysis such as standardization.
245
8
A
ccepted
Table 1: List of
scorpan
environmental covariates used with unit and associated reference
when applicable.
Factor of soil formation Predictor variable
Unit
Reference
Soil
Average soil and sedimentary-deposit
thickness
metre
Landsat Band 3 (red) for year 2014
-
Global Water Table Depth
metre
Landsat Band 4 (NIR) for year 2014
-
Landsat Band 5 (SWIR) for year 2014
-
Landsat Band 7 (SWIR) for year 2014
-
Climate
Long-term averaged mean annual sur-
face temperature (daytime) MODIS
Kelvin
U.S. Land Process archive center
Temperature seasonality
Celsius
Precipitation of driest month
mm
Total annual precipitation
mm
Temperature Annual Range
Celsius
Organisms/vegetation
Global tree cover
%
Cultivated land cover for year 2010
%
Forests cover for year 2010
%
Grasslands cover for year 2010
%
Shrublands cover for year 2010
%
Wetland cover for year 2010
%
Relief
DEM
metre
Terrain slope
radians x 100 -
Multiresolution Index of Valley Bottom
Flatness (MRVBF)
metre x 100
-
SAGA Wetness Index
metre x 10
Landform class: Breaks/Foothills
%
U.S. Geological Survey
Landform class: Flat Plains
%
U.S. Geological Survey
Landform class: High Mountains/Deep
Canyons
%
U.S. Geological Survey
Landform class: Hills
%
U.S. Geological Survey
Landform class: Low Hills
%
U.S. Geological Survey
Landform class: Low Mountains
%
U.S. Geological Survey
Landform class: Smooth Plains
%
U.S. Geological Survey
Parent material/age
Rock type: Acid Plutonics
%
Rock type: Carbonate Sedimentary
%
Rock type: Metamorphics
%
Rock type: Siliciclastic Sedimentary
%
Rock type: Mixed Sedimentary
%
Rock type: Basic Volcanics
%
Rock type: Unconsolidated Sediment
%
Geographical position
X-coordinates
metre
-
Y-coordinates
metre
-
9
A
ccepted
3.2. Practical implementation
The original dataset was randomly split between calibration (80%) and test
(20%) sets (Fig. 1). All six soil properties (clay, silt, sand, OC, pH and N) were
simultaneously selected for either test of calibration. Each soil property was
normalized between 0 and 1 while the covariates were standardized by subtract-
250
ing their mean and dividing by their standard deviation. Recall that categorical
covariates are converted to dummy variables prior the standardization. Two 4-D
matrices are created, one for calibration and one for test, by extracting covari-
ates values in the vicinity of the sampling location. They are of size
n
×
c
×
w
×
h
for the calibration set where
n
is the number of sampling location,
c
= 37
is
255
the number of covariates and
w
=
h
are the dimensions (in number of pix-
els) of the covariates input images. For the test set, the 4-D matrix has size
N
−
n
×
c
×
w
×
h
.
Figure 1: LUCAS sampling locations in France. The squares in blue are the test locations
and in red the calibration locations. Note that a square is used to illustrate the amount of
covariate contextual information included into the model.
Next, a sequential multi-task CNN model is built for predicting jointly the six
soil properties. The CNN model has a common architecture for the convolu-
260
tional and pooling layers followed by separated fully connected layers which
output either one (model for bootstrap) or two (model for variance estimate)
values per soil property. The model specifications are reported in Table 2 and
a graphical representation is given in Fig. 2. Zero padding is always applied
to the convolutional layers to preserve the original size of the input image and
265
conserve information at an early stage of the network.
Soil fraction clay, silt and sand are reported in percent, which must sum to 100.
To handle this compositional constrain, the prediction from the output layer of
10
A
ccepted
Table 2: Layers used in the sequential model built for clay, silt, sand, OC, pH and N prediction.
A graphical representation is given in Fig. 2
Layer type
Shared
Filter size Number of fil-
ters/neurons
Activation
Convolutional
yes
3
×
3
64
ReLU
Max-pooling
yes
2
×
2
-
-
Convolutional
yes
2
×
2
40
ReLU
Dropout (0.1)
yes
-
-
-
Convolutional
yes
2
×
2
22
ReLU
Convolutional
yes
2
×
2
30
ReLU
Dropout (0.1)
yes
-
-
-
Flatten
yes
-
-
-
Fully-connected
no
-
160
ReLU
Dropout (0.3)
no
-
-
-
Fully-connected
no
-
20
ReLU
Dropout (0.2)
no
-
-
-
Fully-connected
no
-
1 or 2
Linear
Concatenate & Softmax clay, silt, sand -
-
-
clay, silt and sand is passed through a softmax layer. A softmax layer takes as
input a vector and returns a vector of the same length, where each value is in
270
the range
(0
,
1)
and the vector adds up to one. The output values of clay, silt
and sand are then multiplied by 100. The softmax function is very similar to
the additive log-ratio transform generally applied to compositional variables in
soil mapping studies.
The CNN model is trained using different window size of image input (
h
=
w
)
275
of 3, 6, 9, 15, 21 and 29 pixels. The optimal window size is chosen based
on the averaged soil properties RMSE of the test set. The parameters were
estimated by minimizing either Eq. 4 or Eq. 9 using the Adam optimizer. It
is of common practice to further separate the calibration set into calibration
(90%) and validation set (10%) before training the model. The calibration set
280
is used to find optimal values of
θ
while the validation set is used to ensure that
the model is not overfitting during the parameter optimization process. Since
the test set is used only for computing the accuracy measures it can not be
used for this purpose. Overfitting is prevented by adding dropout layers which
deactivate neurons of a given layer at random during each batch. A total of 100
285
bootstraps are made so that
B
= 100
. Predictions are made on the centre cells
of a
1
km
×
1
km grid.
Processing was done in R 3.5.1 (R Core Team, 2018), using the keras package
(Allaire & Chollet, 2018) and tensoflow (Abadi et al., 2016) backend. Training
a single model for 500 epochs, a batch size of 350, an input window size of
9
×
9
290
for 37 covariates and 2,357 sampling locations took approximately 2 hours in
parallel on a Linux server 4.4.0-38- generic Ubuntu SMP with 48 cores.
Random forest (RF) is used as a reference model to compare the predictions
11
A
ccepted
Convolutio
nal
Max-pooling
Convolutio
nal
Dropout
Convolutio
nal
Convolutio
nal
Dropout
Fla
�
en
Fully-
connected
Fully-
connected
Fully-
connected
Fully-
connected
Fully-
connected
Fully-
connected
Dropout
Dropout
Dropout
Dropout
Dropout
Dropout
Fully-
connected
Fully-
connected
Fully-
connected
Fully-
connected
Fully-
connected
Fully-
connected
Dropout
Dropout
Dropout
Dropout
Dropout
Dropout
Fully-
connected
Fully-
connected
Fully-
connected
Fully-
connected
Fully-
connected
Fully-
connected
Concatenate
+
Softmax
Shared lay
ers
Clay
Silt
Sand
OC
pH
N
V
ariance
clay
V
ariance
silt
V
ariance
sand
V
ariance
OC
V
ariance
pH
V
ariance
N
Figure
2:
Graphical
represen
tation
of
the
CNN
mo
dels
built
in
this
study
.
Net
w
ork
represen
ted
in
blac
k
is
used
for
th
e
b
o
otstrap
while
net
w
ork
com
bining
blac
k
and
red
color
s
is
used
for
the
v
ariance
estimate.
12
A
ccepted
made by CNN. Random forest has been widely used in DSM studies to model
non-linear relationships between soil properties and environmental covariates
295
(Hengl et al., 2018). For a fair comparison between CNN and RF, the same
calibration and test sets are used for both models. RF forest is trained for each
soil property separately using 1000 trees and fine-tuned parameter values.
4. Results
Figure. 3 shows the RMSE of each soil property for different window size of
300
input images. On average, the RMSE decreases for increasing window size
from
3
×
3
to
9
×
9
pixels. Recall that one pixel has a resolution of 1 km so
that a window size of
9
×
9
pixels includes contextual information surrounding
the soil property measurement up to
(9
×
1)
/
2 = 4
.
5
km. Fig. 3 shows that
using a window size larger than
9
×
9
pixels does not lead to further decrease
305
of the soil properties RMSE, with the exception of pH for which the lowest
RMSE is found using a window size of
15
×
15
pixels. CNN does not allow for
different input window size while predicting using a multi-task model, therefore
all results presented hereafter come from using an input window size of 9
×
9
pixels. The scatterplots of measured against predicted by CNN soil properties
1.0
2.5
5.0
10.0
20.0
3
6
9
15
21
29
Vicinity size
RMSE
Property
Clay
Silt
Sand
OC
pH
N
Figure 3: Effect of the vicinity size of the input image. The RMSE corresponds to the error
between the predictions and measured values in the test set.
310
are shown in Fig. 4. For all soil properties, the predictions generally follow the
1:1 line. The agreement between measured and predicted values of clay and silt
is satisfactory because most values are along the 1:1 line. For sand the predicted
values are more scattered for large values of measured sand. The bulk of OC and
N predictions are gathered on the 1:1 line, with the exception of large measured
315
OC and N values which are strongly underestimated (e.g. measured OC = 200
and predicted OC = 50 g kg
−
1
). Scatterplot of pH shows a different pattern.
13
A
ccepted
OC
pH
N
Clay
Silt
Sand
0
50
100
150
4
6
8
0
5
10
15
20
0
20
40
60
0
25
50
75
0
25
50
75
100
0
25
50
75
100
0
5
10
15
20
0
25
50
75
4
6
8
0
20
40
60
0
50
100
150
Measured
Predicted
Figure 4: Scatterplots of the measured against predicted by CNN soil properties. Clay, silt
and sand are expressed in percent, OC and N in g kg
−
1
.
Prediction of small pH values is more dispersed than prediction of larger ones.
The measured pH values do not exceed 8 while the predictions are often greater
than 8 and sometimes close to 9.
320
CNN predictions are compared to those made by RF (Table 3) as a baseline.
CNN accuracy measures are on average equivalent or better than those of RF.
Clay, silt, sand and pH are better predicted by RF than CNN, as shown by the
r
2
and RMSE. CNN does better for predicting OC and N with a significantly
larger
r
2
for OC (
r
2
for CNN is 0.15 and
r
2
for RF is 0.12). This is confirmed
325
by the CCC values which are greater for OC and N predicted by CNN. While
clay, silt, sand and pH predicted by RF have a larger correlation coefficient with
the measured values than those predicted by CNN, the CCC shows that CNN
predictions are either better or equal to those of RF. Since the CCC assesses the
deviation of the predictions with respect to the 1:1 line it is a useful measure
330
to serve model comparison (Lawrence & Lin, 1989). The ME values show that
prediction are relatively unbiased, with the exception OC predictions which
are either slightly positively of moderately negatively biased for RF and CNN,
respectively. This bias is certainly due to the measured OC values greater than
60 g kg
−
1
that are systematically underestimated. This is visible in Fig. 4.
335
14
A
ccepted
Table 3: Prediction accuracy as evaluated on the test set for CNN and RF.
Soil Property
r
2
RMSE
ME
CCC
RF
CNN
RF
CNN
RF
CNN
RF
CNN
Clay
0
.
31
0
.
22
8
.
90
9
.
87
0
.
32
−
0
.
36
0
.
46
0
.
45
Silt
0
.
44
0
.
40
13
.
50
14
.
38
0
.
93
0
.
55
0
.
60
0
.
62
Sand
0
.
42
0
.
36
16
.
08
17
.
47
−
0
.
57
−
0
.
37
0
.
57
0
.
59
OC
0
.
12
0
.
15
24
.
91
18
.
65
3
.
73
−
6
.
32
0
.
23
0
.
46
pH
0
.
50
0
.
39
0
.
79
0
.
90
0
.
02
0
.
03
0
.
64
0
.
61
N
0
.
20
0
.
24
1
.
54
1
.
45
0
.
22
−
0
.
29
0
.
31
0
.
20
28
6
97
13
28
10
78
0.1
32
12
45
2
Clay
Sand
Silt
¯
0
200
400
600
800
100
km
Figure 5: Maps of the predicted mean (left) and standard deviation (right) for clay, silt and
sand. Values are expressed in percent and sum to 100 for the predictions.
15
A
ccepted
The prediction and prediction standard deviation maps for clay, silt and sand
are presented in Fig. 5 and for OC, pH and N in Fig. 6. For all soil properties,
the maps of the mean in the left-hand side show a smooth but detailed pattern
with significant spatial variation. Recall that prediction of clay, silt and sand are
made with the constraint that they must sum to 100% at any prediction location.
340
The clay, silt and sand maps have different range of predicted values. Variation
of silt content is large, with values ranging from 13 to 97%. In contrast, clay
values do not exceed 45%. Large clay content is found in the East of France,
while the map of silt shows large silt content in the topsoils of the North-
West part of France (Britain included) and a low content in the Massif Central
345
mountains. Sand has its largest proportion in the latter mountains and in the
Landes, while being almost unexistent in the upper North of France. Large
values of OC (> 200 g kg
−
1
) are found in mountainous areas such as in the
Alps, the Massif Central and in the Pyrenees. This is similar to the spatial
pattern of the total Nitrogen (N) which in addition has large values in Britain
350
(> 4 g kg
−
1
). High values of pH (
>
8
), indicating alkaline soils, are found in
most parts of France, with the exceptions of Britain, the Vosges Mountains and
the Massif Central.
16
A
ccepted
64
0.18
12
0
3.7
2.6
5
0
1.4
0.17
400
1.12
N
pH
OC
0
200
400
600
800
100
km
¯
Figure 6: Maps of the predicted mean (left) and standard deviation (right) for OC, pH and
N. Values for OC and N are expressed in g kg
−
1
The maps of the standard deviation in the right-hand side of Fig. 5 and 6 are
17
A
ccepted
Table 4: Evaluation of the uncertainty quantification for the CNN model.
Soil Property
˜
δ
PICP 90%
Clay
0.67 90.51
Silt
0.63 89.49
Sand
0.54 92.88
OC
0.55 79.15
pH
0.68 99.83
N
0.57 98.31
different from those of the mean and seem to not share many common spatial
355
features. They have a lower range of values than the maps of the mean. The
largest value of the standard deviation is about two times smaller for clay,
sand and OC, and about three times smaller for silt, pH and N than their
associated mean maps. For clay, silt, OC and N, largest uncertainty is found
in mountainous areas in the Alps or in the Pyrenees and in Britain for OC.
360
Surprisingly, areas with the lowest uncertainty for silt is found in the North of
France, where the mean map has the largest predicted values. A similar pattern
is observed for sand in the Massif Central where largest sand values have small
standard deviation. For sand, large uncertainty is observed in the Landes area
and in a large patch South of Paris. The pattern of the standard deviation map
365
of pH is rather scaterred, with largest uncertainty found in the North-Western
cost and in the Massif Central.
The median of
δ
in Table 4 shows that the uncertainty is underestimated for
all soil properties. Sand, OC and N are slightly underestimated (
˜
δ
< 0.57)
while clay, silt and pH are moderately underestimated (
˜
δ
< 0.68). The 90%
370
PICP shows however that the 90% prediction interval covers satisfactorily the
observed values of clay, silt and sand but is too wide for pH and N and slightly
too narrow for OC.
18
A
0.00
0.25
0.50
0.75
1.00
0.00
0.25
0.50
0.75
1.00
Probability q
Proportion in this interv
al
Property
Clay
Silt
Sand
OC
pH
N
Figure 7: Accuracy plot of the soil properties.
This is confirmed by the accuracy plot in Fig. 7: prediction intervals as obtained
for the soil properties are too narrow, leading to an overall underestimation of
375
the uncertainty. This is more severe for pH where deviation from the 1:1 line
is entirely due to underestimation of the uncertainty. Uncertainty for OC and
N is underestimated for small values of nominal
q
while it is overestimated for
nominal
q
values in the range
0
.
6
−
1
.
5. Discussion
380
Effect of the input window size
The window size of the input images had a significant impact on prediction
accuracy, as shown by the soil properties RMSE on the test set. This is an
expected result also reported in Wadoux et al. (2019) and Padarian et al. (2019).
The window size closely relates to the amount of contextual information supplied
385
to the model. CNN integrates spatial context by accounting for covariate pixels
in the neighbour of a sampling location. More extra context improves the model
predictions, but a too large amount of contextual information certainly acts as
noise (Padarian et al., 2019). This confirms the study made by Smith et al.
(2006) who found that there is an optimal range of window size in a case study
390
deriving terrain attributes from a DEM for use in a soil survey. In our case
study mapping soil properties at national scale, a window size between
9
×
9
pixels and
15
×
15
pixels was found optimal. Outside this range the prediction
can be as much as 15% less accurate than the highest accuracy values. This
is because soil property spatial variation is governed by complex relationships
395
with soil forming factors. For example if the soil forming factor (i.e. one of
19
A
ccepted
the covariate) is homogeneous, one would like to increase the window size so
as to capture sufficient information for the modelling. It is technically possible
to integrate a covariate dependent window size in CNN, but it is not obvious
that the improvement of the prediction outweighs the substantial increase of
400
computational complexity.
A window size from
9
×
9
to
15
×
15
pixels is similar to incorporating contextual
information in a radius of a location up to 4.5 to 7.5 km. Several authors
(e.g. Wadoux et al. (2019), Padarian et al. (2019)) have suggested that the
window size could be associated to the range of spatial autocorrelation of the soil
405
property. Wadoux et al. (2019) reported a range of a fitted spherical variogram
of 329 m and 275 m for top- and subsoil organic carbon mapping. This was close
to their optimal window size radius between 260 and 360 m. Padarian et al.
(2019) compared their optimal window size to the variogram range from other
studies mapping the same soil property in similar conditions. The authors found
410
similar values of autocorrelation range between 150 and 450 m. This hypothesis
was verified by fitting exponential function to the sample variograms of the soil
properties. I found values of the fitted range parameter between 3 km for OC
and up to 8 km for sand. While there is large variability, this is close to the
radius of the window size that is found optimal. However it is not possible to
415
draw conclusions. Investigating this matter would certainly make a valuable
extension for future DSM studies.
Prediction accuracy
For this case study, CNN provided good predictive ability as shown by the
RMSE, ME,
r
2
and CCC. Compared to RF as a baseline, CNN performs equally
420
on average, while being more accurate for mapping OC and silt. Recent studies
using deep learning for soil mapping (e.g. Behrens et al. (2018)) have concluded
that deep learning outperforms RF. This is here clearly not the case. However,
those studies were using a small number of covariates to train the model. Padar-
ian et al. (2019) used a DEM with two of its terrain derivatives in addition to
425
rainfall and precipitation while Wadoux et al. (2019) used solely a DEM, a map
of NDVI and a Landsat ETM band 5 image. A reason for CNN to perform well
with a small number of covariate is possibly that the model creates a large num-
ber of hyper-covariates from the original images during convolution, while RF is
purely data-driven and empirical on the existing covariates. When the number
430
of covariate is large, this effect becomes negligible. In fact, it is acknowledged
that RF is clearly favoured by a large set of covariates (Nussbaum et al., 2018),
while being rapidly limited if the number of covariates is too small. However the
present study is the first to employ a large number of environmental covariates
for soil mapping using deep learning. The impact of the number of covariates
435
on model predictions accuracy has to be further investigated.
20
ccepted
Interpretation of the map features
It is beyond the scope of this study to give an interpretation of all the map
features, but I summarize the most striking ones. The soil texture seems to be
much influenced by the geology, as already noted by Ballabio et al. (2016). Clay
440
content is high (> 40%) for a large part of Eastern France dominated by early
Cretaceous lime and sandstone and middle or late Jurassic limestone. Silt has
very high proportion in luvisols on late Cretaceous limestone and on cambisols
in Britain. The pattern of sand content is close to that of old massifs such in
the Massif Central of the Vosges characterized by late carboniferous metamor-
445
phic bedrocks. In addition, sand content is high in the Landes, constituted of
Pliocene fluviatil rocks. The maps of clay, silt and sand closely resemble those
made by others using the same dataset (e.g. Ballabio et al. (2016)). Predicted
maps also share the same pattern than SoilGrid products for France (Hengl
et al., 2014). The maps of OC, pH and N underline the influence of climate,
450
landuse and geology. OC follows the same distribution as the precipitation and
temperature with larger OC content in mountainous areas. In Britain, this is
due to slurry production related C input (Meersmans et al., 2012). The overall
pattern is in accordance with Meersmans et al. (2012) and Martin et al. (2010).
Properties pH and N are more difficult to interpret, showing strong correlation
455
with precipitation and temperature, but also geology. PH seems to be acidic in
areas of metamorphic and plutonite rocks, while N content is high in all moun-
tainous areas, with smaller temperature and precipitation rate. As for the soil
texture, OC and pH maps are close to those produced by SoilGrid (Hengl et al.,
460
Uncertainty quantification
It is the first study to propose a method to quantify the uncertainty of the pre-
diction for mapping using a neural network model. On average, the prediction
uncertainty was slightly underestimated (Table 4). This is an expected result
reported in previous studies (Khosravi et al., 2011). The reason is that estima-
465
tion of the
σ
ˆ
f
is dependent on
B
neural network. Each network is calibrated
using a bootstrap sample of the input data. A specific sample might be unrep-
resentative of the population and cause inaccurate estimation of the variance
parameter and underestimation of the total variance. Other methods for neural
networks uncertainty quantification exist, such as the Delta (De Vleaux et al.,
470
1998), or the Bayesian uncertainty analysis (MacKay, 1992) method, but those
have not been tested. Albeit underestimated, the uncertainty of the prediction
was accurately quantified, as shown by the accuracy plots and PICP. Caubet
et al. (2019) reported the PICP of the LUCAS prediction (from the maps made
by Ballabio et al. (2016)) of clay and sand for France and found values of 8
475
and 15%, respectively. In the present study were obtained 90.51% for clay and
92.88% for sand, which are values very close to the expected value of proba-
bility interval. The uncertainty was therefore much better quantified than this
existing work using the same dataset in similar conditions.
21
A
ccepted
Multivariate mapping using CNN
480
Six soil properties were predicted using a single multi-task CNN model. The
CNN shared a common architecture for all soil properties. This reduces both
the risk of overfitting and computational resources that would be needed to fit
each model separately. This is important for future DSM studies because the
number of available geodata is constantly increasing (Nussbaum et al., 2018).
485
In addition, the multi-task CNN architecture can easily be modified to predict
at several depths. Wadoux et al. (2019) and Padarian et al. (2019) have shown
that this is feasible and lead to a substantial increase of prediction accuracy
in deeper soil layers. Several attempts have been made for multivariate soil
mapping (e.g. by Angelini et al. (2017)). However, there have been very little,
490
if any, interest for multivariate soil mapping using machine learning techniques.
Hengl et al. (2018) was the first to investigate the use of a multi-output random
forest. However, the method proposed by Hengl et al. (2018) has the major
disadvantage that the data size increases rapidly as the number of outputs
expands. In addition, deriving prediction intervals for each output separately
495
remains a challenge. In the case study presented here, it is shown that building
a single model for multiple outputs is feasible, and lead to prediction accuracy
comparable to those made by a univariate RF model.
Correlation between outputs
Recent studies (e.g. Angelini et al., 2017) argued that one of the strengths of a
500
multivariate model is its capability to preserve the correlation between soil prop-
erties. In Angelini et al. (2017), the correlation between properties is assessed as
resulting from a calibrated SEM model. One cannot assess internal correlation
between properties in a ML model, but it is possible to measure whether the
correlation between original and predicted soil properties is preserved. In this
505
study, the correlation between predicted outputs is on average slightly better
preserved by the univariate random forest model than by the CNN model, as
assessed by the properties Pearson’s
r
correlation coefficient matrix. This can
be explained by the
r
2
values of the predictions made by the RF model that are,
for most properties, closer to one than those of the CNN model. However, since
510
the correlation between outputs is not modelled explicitly, it would be mislead-
ing to conclude that the model retains the correlation between outputs simply
by assessing the correlation or covariation between its predictions. I hope this
clarifies why the correlation between outputs was not used for interpretation
purposes in this study. One can include explicitly the correlation between out-
515
puts by calibrating additional stochastic variables together with the CNN model
parameters (Uria et al., 2016). Another solution is to modify the loss function
so that a criterion related to the absolute difference between the correlation
among original and predicted soil properties is minimized jointly with the MSE.
Note also that in this study, a CNN with a shared architecture is built, and
520
with separated branches for each soil property. I speculate that the correlation
between outputs would be better retained by using a shared architecture only,
22
A
ccepted
and by predicting directly after the flattening operation. The effect of the neural
network architecture on the predictions and methods to include explicitly cor-
relation between outputs need further investigation. I welcome more research
525
in this area.
Assumptions made during modelling
The method described in this study for uncertainty quantification assumes inde-
pendent and normally distributed residuals. The latter assumption was tested
by visual inspection of the residuals and computation of a quantile-quantile (Q-
530
Q) plot (not shown). It was found that the Gaussian assumption of the resid-
uals was satisfied. In cases where the Gaussian assumption is too restrictive,
one could use a transformation of the data prior to modelling (e.g. logarithm,
square-root or Box-Cox transform). Remaining spatial autocorrelation in the
residuals was further tested to ensure that the assumption of independence was
535
satisfied. Spatial autocorrelation in the residuals indicates that the prediction
might be biased, and that the quantified uncertainty is possibly underestimated.
The Moran’s I (MI) (Bivand & Wong, 2018) was used to test for spatial auto-
correlation of the residuals of the maps presented in Fig. 5 and 6. The values
of the MI range from -1 to 1, with 0 indicating no autocorrelation. Ten nearest
540
neighbours were used to compute the MI on the original values and residuals
of the soil properties. Results indicate significant spatial autocorrelation in the
original soil properties, with a MI value larger than 0.27 for all soil properties
(e.g. MI value for pH is 0.41), while being smaller than 0.1 on the residuals
(MI value for residuals of pH is 0.05). Thus, the CNN model was efficient to
545
eliminate spatial autocorrelation. In case there would remain autocorrelation
in the residuals, predictions and prediction uncertainty quantification would be
improved by kriging the residuals, and by summing the kriged mean and vari-
ance maps to the predicted mean and variance maps made by the CNN model.
An example can be found in Rossel et al. (2015).
550
6. Conclusion
From the results and discussion I draw the following conclusions:
• It is feasible to use convolutional neural network to map soil properties
simultaneously. Extensions can further be made to predict at multiple
depths. CNN has the advantage that it incorporates explicitly the con-
555
textual information of the covariates in the neighbour of a location. In
addition, CNN as for other machine learning models, does not rely on rigid
statistical assumptions about the distribution of the soil properties.
• The window size of the input covariate images had a significant impact on
prediction accuracy. The optimal window size closely relates to the range
560
of the spatial autocorrelation of the soil property. However, there is need
23
A
ccepted
for further research so as to generate rules and enable to chose
a priori
the optimal size.
• The predicted maps of clay, silt, sand, OC, pH and N have a detailed
pattern with significant spatial variation. These maps showed great simi-
565
larities with those of same soil properties produced by others using differ-
ent mapping techniques. Statistical validation of the prediction accuracy
indicated that CNN performs on average equally to random forest.
• Bootstrap plus variance estimate was used to quantify the uncertainty of
the predictions. Each soil property has its own standard deviation map,
570
which can be used to derive prediction intervals. Validation of the uncer-
tainty assessment indicated that the uncertainty is accurately quantified,
albeit slightly underestimated. Comparison with other studies using the
same dataset over France showed that the proposed method quantifies
much better the uncertainty.
575
Acknowledgements
Alexandre Wadoux received funding from the European Union’s Seventh Frame-
work Programme for research, technological development and demonstration
under grant agreement no 607000. The author thank ISRIC - World Soil In-
formation and Tom Hengl, Envirometrix Ltd, Wageningen, for providing the
580
covariates used in the case study.
References
Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M.,
Ghemawat, S., Irving, G., Isard, M. et al. (2016). Tensorflow: a system for
large-scale machine learning. In
OSDI
(pp. 265–283). volume 16.
585
Akpa, S. I. C., Odeh, I. O. A., Bishop, T. F. A., & Hartemink, A. E. (2014).
Digital mapping of soil particle-size fractions for Nigeria.
Soil Science Society
of America Journal
,
78
, 1953–1966.
Allaire, J. J., & Chollet, F. (2018).
keras: R Interface to ’Keras’
. URL:
//CRAN.R-project.org/package=keras
r package version 2.2.0.
590
Angelini, M. E., Heuvelink, G. B. M., & Kempen, B. (2017). Multivariate
mapping of soil with structural equation modelling.
European Journal of Soil
Science
,
68
, 575–591.
Ballabio, C., Panagos, P., & Monatanarella, L. (2016). Mapping topsoil physical
properties at European scale using the LUCAS database.
Geoderma
,
261
,
595
110–123.
24
A
ccepted
Behrens, T., Schmidt, K., MacMillan, R. A., & Viscarra Rossel, R. A. (2018).
Multi-scale digital soil mapping with deep learning.
Scientific reports
,
8
,
15244–15244.
Bivand, R. S., & Wong, D. W. S. (2018). Comparing implementations of global
600
and local indicators of spatial association.
TEST
,
27
, 716–748.
Caubet, M., Dobarco, M. R., Arrouays, D., Minasny, B., & Saby, N. P. (2019).
Merging country, continental and global predictions of soil texture: Lessons
from ensemble modelling in France.
Geoderma
,
337
, 99–110.
Chen, J., Chen, J., Liao, A., Cao, X., Chen, L., Chen, X., He, C., Han, G.,
605
Peng, S., Lu, M. et al. (2015). Global land cover mapping at 30 m resolution:
A pok-based operational approach.
ISPRS Journal of Photogrammetry and
Remote Sensing
,
103
, 7–27.
De Vleaux, R. D., Schumi, J., Schweinsberg, J., & Ungar, L. H. (1998). Pre-
diction intervals for neural networks via nonlinear regression.
Technometrics
,
610
40
, 273–282.
Deutsch, C. V. (1997). Direct assessment of local accuracy and precision.
Geo-
statistics Wollongong
,
96
, 115–125.
Efron, B., & Tibshirani, R. J. (1994).
An Introduction to the Bootstrap
vol-
ume 57. CRC press, New York.
615
Fan, Y., Li, H., & Miguez-Macho, G. (2013). Global patterns of groundwater
table depth.
Science
,
339
, 940–943.
Goodfellow, I., Bengio, Y., Courville, A., & Bengio, Y. (2016).
Deep Learning
volume 1. MIT press Cambridge.
Goovaerts, P. (1997).
Geostatistics for Natural Resources Evaluation
. Oxford
620
Univ. Press, New York.
Hansen, M. C., Potapov, P. V., Moore, R., Hancher, M., Turubanova, S. A. A.,
Tyukavina, A., Thau, D., Stehman, S. V., Goetz, S. J., Loveland, T. R.
et al. (2013). High-resolution global maps of 21st-century forest cover change.
Science
,
342
, 850–853.
625
Hartmann, J., & Moosdorf, N. (2012). The new global lithological map database
GLiM: A representation of rock properties at the Earth surface.
Geochemistry,
Geophysics, Geosystems
,
13
, 1–37.
Hengl, T., de Jesus, J. M., MacMillan, R. A., Batjes, N. H., Heuvelink, G.
B. M., Ribeiro, E., Samuel-Rosa, A., Kempen, B., Leenaars, J. G. B., Walsh,
630
M. G. et al. (2014). Soilgrids1km-global soil information based on automated
mapping.
PloS one
,
9
, e105992.
25
A
ccepted
Hengl, T., Nussbaum, M., Wright, M. N., Heuvelink, G. B. M., & Gräler, B.
(2018). Random forest as a generic framework for predictive modeling of
spatial and spatio-temporal variables.
PeerJ
,
6
, e5518.
635
Heuvelink, G. B. M. (2014). Uncertainty quantification of globalsoilmap prod-
ucts. In
GlobalSoilMap: basis of the global spatial soil information system.
Proceedings of 1st GlobalSoilMap Conference
(pp. 335–340).
Heuvelink, G. B. M., Kros, J., Reinds, G. J., & De Vries, W. (2016). Geosta-
tistical prediction and simulation of european soil property maps.
Geoderma
640
Regional
,
7
, 201–215.
Jones, A., Montanarella, L., Jones, R. et al. (2005).
Soil atlas of Europe
. Euro-
pean Commission.
Karger, D. N., Conrad, O., Böhner, J., Kawohl, T., Kreft, H., Soria-Auza,
R. W., Zimmermann, N. E., Linder, H. P., & Kessler, M. (2017). Climatologies
645
at high resolution for the earth’s land surface areas.
Scientific data
,
4
, 170122.
Khosravi, A., Nahavandi, S., Creighton, D., & Atiya, A. F. (2011). Comprehen-
sive review of neural network-based prediction intervals and new advances.
IEEE Transactions on neural networks
,
22
, 1341–1356.
Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization.
650
3
rd
International Conference for Learning Representations, San Diego, 2015
,
(pp. 1–15).
Lakshminarayanan, B., Pritzel, A., & Blundell, C. (2017). Simple and scalable
predictive uncertainty estimation using deep ensembles. In
Advances in Neural
Information Processing Systems
(pp. 6402–6413).
655
Lark, R. M. (2000). A comparison of some robust estimators of the variogram
for use in soil survey.
European Journal of soil science
,
51
, 137–157.
Lark, R. M., Milne, A. E., Addiscott, T. M., Goulding, K. W. T., Webster, C. P.,
& O’Flaherty, S. (2004). Scale-and location-dependent correlation of nitrous
oxide emissions with soil properties: an analysis using wavelets.
European
660
Journal of Soil Science
,
55
, 611–627.
Lawrence, I., & Lin, K. (1989). A concordance correlation coefficient to evaluate
reproducibility.
Biometrics
, (pp. 255–268).
LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard,
W., & Jackel, L. D. (1989). Backpropagation applied to handwritten zip code
665
recognition.
Neural computation
,
1
, 541–551.
MacKay, D. J. C. (1992). The evidence framework applied to classification
networks.
Neural computation
,
4
, 720–736.
26
A
ccepted
Martin, M. P., Wattenbach, M., Smith, P., Meersmans, J., Jolivet, C., Boulonne,
L., & Arrouays, D. (2010). Spatial distribution of soil organic carbon stocks
670
in France: Discussion paper.
Biogeosciences
,
8
, 1053–1065.
McBratney, A. B., Santos, M. L. M., & Minasny, B. (2003). On digital soil
mapping.
Geoderma
,
117
, 3–52.
Meersmans, J., Martin, M. P., Lacarce, E., De Baets, S., Jolivet, C., Boulonne,
L., Lehmann, S., Saby, N. P. A., Bispo, A., & Arrouays, D. (2012). A high
675
resolution map of French soil organic carbon.
Agronomy for Sustainable De-
velopment
,
32
, 841–851.
Miller, B. A., Koszinski, S., Wehrhan, M., & Sommer, M. (2015). Impact of
multi-scale predictor selection for modeling soil properties.
Geoderma
,
239
,
97–106.
680
Nix, D. A., & Weigend, A. S. (1994). Estimating the mean and variance of the
target probability distribution. In
Neural Networks, IEEE World Congress
on Computational Intelligence
(pp. 55–60). IEEE volume 1.
Nussbaum, M., Spiess, K., Baltensweiler, A., Grob, U., Keller, A., Greiner, L.,
Schaepman, M. E., & Papritz, A. (2018). Evaluation of digital soil mapping
685
approaches with large sets of environmental covariates.
SOIL
,
4
, 1–22.
Olaya, V., & Conrad, O. (2009). Geomorphometry in SAGA.
Developments in
Soil Science
,
33
, 293–308.
Padarian, J., Minasny, B., & McBratney, A. B. (2019). Using deep learning for
digital soil mapping.
SOIL
,
5
, 79–89.
690
Pelletier, J. D., Broxton, P. D., Hazenberg, P., Zeng, X., Troch, P. A., Niu, G.-
Y., Williams, Z., Brunke, M. A., & Gochis, D. (2016). A gridded global data
set of soil, intact regolith, and sedimentary deposit thicknesses for regional
and global land surface modeling.
Journal of Advances in Modeling Earth
Systems
,
8
, 41–65.
695
R Core Team (2018).
R: A Language and Environment for Statistical Com-
puting
. R Foundation for Statistical Computing Vienna, Austria. URL:
Robinson, N., Regetz, J., & Guralnick, R. P. (2014). EarthEnv-DEM90: A
nearly-global, void-free, multi-scale smoothed, 90m digital elevation model
700
from fused ASTER and SRTM data.
ISPRS Journal of Photogrammetry and
Remote Sensing
,
87
, 57–67.
Rossel, R. A. V., Chen, C., Grundy, M. J., Searle, R., Clifford, D., & Camp-
bell, P. H. (2015). The Australian three-dimensional soil grid: AustraliaâĂŹs
contribution to the GlobalSoilMap project.
Soil Research
,
53
, 845–864.
705
27
A
ccepted
Shrestha, D. L., & Solomatine, D. P. (2008). Data-driven approaches for es-
timating uncertainty in rainfall-runoff modelling.
International Journal of
River Basin Management
,
6
, 109–122.
Smith, M. P., Zhu, A.-X., Burt, J. E., & Stiles, C. (2006). The effects of DEM
resolution and neighborhood size on digital soil survey.
Geoderma
,
137
, 58–
710
69.
Tóth, G., Jones, A., & Montanarella, L. (2013). The LUCAS topsoil database
and derived information on the regional variability of cropland topsoil prop-
erties in the European Union.
Environmental Monitoring and Assessment
,
185
, 7409–7425.
715
Tucker, C. J., Grant, D. M., & Dykstra, J. D. (2004). NASA’s global orthorec-
tified landsat data set.
Photogrammetric Engineering & Remote Sensing
,
70
,
313–322.
Uria, B., Côté, M.-A., Gregor, K., Murray, I., & Larochelle, H. (2016). Neu-
ral autoregressive distribution estimation.
The Journal of Machine Learning
720
Research
,
17
, 7184–7220.
Wadoux, A. M. J. C., Brus, D. J., & Heuvelink, G. B. M. (2018). Account-
ing for non-stationary variance in geostatistical mapping of soil properties.
Geoderma
,
324
, 138–147.
Wadoux, A. M. J. C., Padarian, J., & Minasny, B. (2019). Multi-source data
725
integration for soil mapping using deep learning.
SOIL
,
5
, 107–119.
Xu, S., An, X., Qiao, X., Zhu, L., & Li, L. (2013). Multi-output least-squares
support vector regression machines.
Pattern Recognition Letters
,
34
, 1078–
1084.
28