Lawrence Berkeley National Laboratory
LBL Publications
Title
Regional-scale soil carbon predictions can be enhanced by transferring global-scale
soil–environment relationships
Permalink
https://escholarship.org/uc/item/05j3c4qt
Journal
Geoderma, 461
ISSN
0016-7061
Authors
Zhang, Lei
Yang, Lin
Ma, Yuxin
et al.
Publication Date
2025-09-01
DOI
10.1016/j.geoderma.2025.117466
Copyright Information
This work is made available under the terms of a Creative Commons Attribution
License, available at https://creativecommons.org/licenses/by/4.0/
Peer reviewed
Powered by the California Digital Library
University of California
Regional-scale soil carbon predictions can be enhanced by transferring
global-scale soil
–
environment relationships
Lei Zhang
,
, Lin Yang
, Yuxin Ma
, A-Xing Zhu
, Ren Wei
, Jie Liu
, Mogens H. Greve
,
Chenghu Zhou
,
a
School of Geography and Ocean Science, Nanjing University, Nanjing, China
b
Climate and Ecosystem Science Division, Lawrence Berkeley National Laboratory, Berkeley, CA, USA
c
New South Wales Department of Climate Change, Energy, the Environment and Water, Parramatta, NSW, Australia
d
Department of Geography, University of Wisconsin-Madison, Madison, WI, USA
e
Department of Agroecology, Aarhus University, Tjele, Denmark
f
State Key Laboratory of Resources and Environmental Information System, Institute of Geographical Sciences and Natural Resources Research, Chinese Academy of
Sciences, Beijing, China
A R T I C L E I N F O
Keywords:
Soil organic carbon
Soil
–
environment relationships
Soil mapping
Pre-trained model
Model transferability
Deep learning
Global and regional scales
A B S T R A C T
Accurate modelling and mapping soil organic carbon are crucial for supporting soil health restoration and
climate change mitigation at both regional and global scales. However, regional soil predictions often suffer from
data scarcity and high prediction uncertainty. Utilizing a pre-trained global-to-regional soil carbon predictive
model can be a potential solution to address this challenge. Despite its promise, how to construct and apply the
global-scale model to enhance regional-scale soil carbon mapping remains largely unexplored. Here, we propose
the Global Soil Carbon Pre-trained Model (GSoilCPM), a deep-learning-based domain adaptative model, to
enhance regional-scale soil carbon predictions. Based on large amount of environmental covariate data and
106,167 soil samples across the globe, we verify our hypothesis of the effectiveness of this
’
global-to-regional
’
modelling strategy. The pre-trained model can be then transferred and fine-tuned to bridge the regional- and
global-scale soil
–
environment relationships. We applied and validated this modelling strategy in four regional-
scale study areas, three in the Northern Hemisphere and one in the Southern Hemisphere, each with distinct
environmental background. Compared to traditional modelling approaches as a baseline, four case studies all
demonstrated significant improvement in prediction accuracy across diverse environments and varying data
availabilities. The average percentage improvement across all regions is 10.93% (absolute values decreased by
1.20 g kg
−
1
averagely) in MAE and 29.04% (absolute values increased by 0.10 averagely) in CCC. The appli-
cability and future horizons of using GSoilCPM were further discussed. We further reveal that regions with fewer
soil samples or lower baseline accuracy benefit more from the pre-trained global model. Our findings highlight
the advantages of leveraging the generalized knowledge from global models to enhance specifically localized soil
modelling, positioning a potential paradigm shift in digital soil mapping, and far-reaching implications for soil
monitoring and land management.
1. Introduction
Soil organic carbon (SOC), a critical component of the Earth
’
s carbon
cycle, plays a fundamental role in maintaining soil fertility, regulating
greenhouse gas emissions, and supporting overall ecosystem health
(
Tiessen et al., 1994; Schmidt et al., 2011; Sanderman et al., 2017
).
Accurate predicting and mapping SOC are essential for understanding
carbon dynamics at both regional and global scales, guiding sustainable
land management practices, and contributing to climate change miti-
gation strategies (
Sanchez et al., 2009; Chen et al., 2022; Huang et al.,
). Nevertheless, our current understanding of
SOC distributions is frequently impeded by high uncertainties stemming
from limited sample data and the underdevelopment of modelling
techniques (
Traditional soil mapping methods rely on experienced soil surveyors
who are familiar with the local area and spend considerable time in field
* Corresponding author.
E-mail addresses:
(L. Zhang),
(L. Yang).
Contents lists available at
Geoderma
journal homepage:
www.elsevier.com/locate/geoderma
https://doi.org/10.1016/j.geoderma.2025.117466
Received 9 April 2025; Received in revised form 19 June 2025; Accepted 29 July 2025
Available
online
5
August
2025
0016-7061/©
2025
The
Author(s).
Published
by
Elsevier
B.V.
This
is
an
open
access
article
under
the
CC
BY-NC-ND
license
work, manually creating maps using the polygon data model (lines,
points, and polygon shapes) (
Bliss et al., 1995; Minasny and McBratney,
). Recent advancements in geostatistics
(
Heuvelink and Webster, 2001; Hengl et al., 2004; Heuvelink et al.,
), machine learning algorithms (
Wadoux et al., 2020; Meng et al., 2022, 2024a; Helfenstein et al., 2024;
Guo et al., 2025; Zhang et al., 2025
), and increasingly accessible land
surface datasets (e.g., geographical variables derived from remote
sensing observations) (
Mulder et al., 2011; Ivushkin et al., 2019; He
) have introduced alternative ways to predict and map soil
information with less manual labor. Data-driven approaches facilitated
us to build soil
–
environment relationships from soil samples and envi-
ronmental covariates, and then to apply these fitted relationships to
generate predicted soil maps. That is where digital soil mapping (DSM)
methods come into play (
McBratney et al., 2003; Minasny and
McBratney, 2016; Ma et al., 2019; Liu et al., 2022
). DSM technique has
been widely applied to soil carbon mapping and has yielded satisfactory
results at both global and regional scales (
et al., 2022; Wang et al., 2023; Meng et al., 2024b; Zhang et al., 2022a,
). Despite these advancements, we are still facing a ‘data
hungry
’
problem in soil mapping field, especially when the soil data are
limited in specific regions (
Zhu et al., 2015; Zhang et al., 2021; Cui et al.,
). The scarcity of soil observations makes it difficult to construct
robust soil
–
environment relationships, which hampers the ability to
accurately capture soil carbon patterns at regional scales.
A potential strategy to overcome this bottleneck is to construct a pre-
trained domain adaptive model trained on globally available soil and
environmental data. This model would operate as a ‘knowledge
’
on
soil
–
environment relationships across the globe, and could then be
adapted (e.g., fine-tuned or transferred) for the regional-scale SOC
predictions. Given the success of pre-trained and transfer modelling
strategy in fields outside of soil mapping (
Awais et al., 2023; Moor et al.,
2023; Wu et al., 2023; Hong et al., 2024
), it is reasonable to explore
whether such an approach could be applied for soil carbon predictions
and whether it could help alleviate the issue of insufficient soil obser-
vations at regional scales and improve mapping accuracy. While a
similar ‘global-to-local
’
approach has been used in soil spectroscopic
modelling (
Shen et al., 2022; Viscarra Rossel et al., 2024
), a clear
research gap remains in applying this concept and designing appropriate
modelling frameworks for SOC mapping.
Another motivation for this study lies in the fact that the demand for
accurate regional-scale soil predictions is often difficult to directly fulfill
using existing global-scale soil map products. Some efforts have pro-
duced global maps of soil carbon at medium spatial resolutions, such as
the well-known Harmonized World Soil Database (HWSD) (
)
and SoilGrids product (
Hengl et al., 2014, 2017; Poggio et al., 2021
).
However, recent studies have highlighted significant uncertainties in
these global soil gridded datasets (
Dai et al., 2019; Lilburne et al., 2024
).
This is largely due to incomplete coverage of global soil profiles, limiting
representation across all areas (
). Moreover, global-
scale modelling often underestimates or biases the high spatial vari-
ability of soil carbon at local scales. This is because spatial predictions
tend to smooth distribution tails (
) and encounter
gaps in predictor space where training sample data is insufficient (
). In addition, models developed for global applica-
tions, such as the random forest (RF) models for producing SoilGrids,
cannot be directly transferred to local areas. These factors impede the
potential re-utilization of global-scale data and models, which would be
a valuable foundational resource for enhancing regional-scale SOC
predictions.
To address this challenge, we propose leveraging a Global Soil Car-
bon Pre-trained Model (GSoilCPM), initially trained on global-scale soil
profiles with SOC observations and remote sensing derived environ-
mental covariate datasets, to enhance regional-scale SOC mapping. The
model architecture not only adopts soil formation theory (
) to differently process the environmental
covariates influencing soil by variable category, but also uses deep
convolutional networks to separately extract latent feature from each
covariate category with varying window sizes that captures spatial
contextual information. The pre-trained model can then be fine-tuned
with an additional round of parameter optimization to perform
regional SOC mapping tasks.
Here, we aim to determine whether the proposed deep-learning (DL)-
based model can achieve competitive performance compared to the
widely used ensemble machine learning models at the global scale.
Then, we hypothesize that the challenges of data scarcity and complex
nonlinear patterns in a target region can be mitigated by utilizing the
pre-trained global-level model. To investigate this hypothesis, we
applied a transfer learning approach to bridge the gap between global
and regional soil
–
environment relationships. This allows the generalized
knowledge captured by the global model to be localized, enhancing
regional SOC modelling. Establishing this ‘global-to-regional
’
modelling
strategy is grounded in the idea that, while local soil conditions are
shaped by site-specific processes, many underlying mechanisms driving
soil variations are governed by universal geographical and ecological
principles. Therefore, those macro-scale regularities captured in a global
model can serve as a knowledge prior and can provide a structured
starting point for guiding regional soil predictions.
In this study, we first introduced the GSoilCPM model and trained it
using the latest global soil profile database from the World Soil Infor-
mation Service (WoSIS), integrated with various environmental cova-
riates, including reflectance bands from satellite observations, climate,
topography, vegetation and parent materials. We then applied the
GSoilCPM model to generate SOC maps in four regions with four distinct
landscape backgrounds and varying data availabilities. These results
were compared with predictions from SoilGrids, RF models, and DL
models without pre-training. Finally, we examined the advantages of
using the global-scale model to enhance regional soil carbon modelling,
supported by varying sizes of sample data in each case study areas,
thereby illustrating the applicability and prospects of this modelling
strategy.
2. Materials and methods
2.1. Global-scale datasets
2.1.1. Soil samples
The global-scale soil sample data were sourced from the WoSIS
database (
), one of the largest and most
comprehensive repositories of harmonized soil profile data worldwide.
For this study, we used the
“
WoSIS snapshot
–
September 2019
”
, which
comprises 196,498 geo-referenced soil profiles worldwide with over
832,000 soil layers or horizons. We filtered the profiles to include only
those with measurements of SOC content (g kg
−
1
), and then checked for
the layer observations up to a depth of 0.3 m. The profiles with low
accuracy of the geographical coordinates (i.e., missing the information
of degree, minute or second data) and those flagged as containing sur-
face litter were excluded. A total number of 106,167 global soil samples
were finally obtained for this study. To estimate the mean SOC content
in topsoil (0
–
0.3 m), we employed the equal-area spline algorithm to fit
multiple SOC observations at different depth intervals. This method has
been proven to be superior to alternative functions (
). The average of the fitted SOC values from 0 to 0.3
m was calculated as SOC content in topsoil for each sample location.
After above data processing, the global soil profile locations, along with
their topsoil organic carbon content values are shown in
a
. These
profiles provide a broad geographic coverage, ensuring a representative
sample set of global soil distributions across different environmental
conditions.
2.1.2. Environmental covariates that related with SOC
The environmental covariates contain a series of variables that
L. Zhang et al.
Geoderma
461
(2025)
117466
2
influence soil carbon variations (
). As one of the aims of this
study is to build a pre-trained large-scale model that links SOC with
environmental conditions, we selected most basic environmental vari-
ables as covariates, according to the concepts of soil formation factors
and soil
–
landscape relationships (
Jenny, 1941; Hudson, 1992; Zhu et al.,
2001; McBratney et al., 2003; Zhu et al., 2018
). These covariates include
climate, topography, vegetation, and parent material. We also collected
remote-sensing-based surface reflectance information with seven spec-
tral bands from the moderate-resolution imaging spectroradiometer
(MODIS) data product. Details of these covariates are shown in
, and are also described as follows.
The remote sensing images that include multiple reflectance bands
can be used as important predictors for soil mapping. Different soil
properties, such as soil texture, moisture content, organic matter and
mineral composition, reflect and absorb light differently across various
wavelengths (
Viscarra Rossel et al., 2006a, 2006b; Mulder et al., 2011
).
Thus, it is useful to feed the spectral bands into the model for analyzing
reflectance patterns and inferring soil characteristics. We collected the
reflectance data with seven bands (620
–
670 nm, 841
–
876 nm, 459
–
479
nm, 545
–
565 nm,1230
–
1250 nm, 1628
–
1652 nm, and 2105
–
2155 nm)
from the Moderate Resolution Imaging Spectroradiometer (MODIS)
MCD43A4 (version 6.1) data product (
). This
dataset provides Nadir Bidirectional Reflectance Distribution Function
(BRDF)-Adjusted Reflectance (NBAR) values, with a 500 m spatial
resolution.
The temperature and precipitation variables obtained from World-
Clim (version 2) (
) were used to represent the
climate condition for each sample location. In addition to the annual
mean temperature and annual precipitation, the variables of tempera-
ture seasonality and precipitation seasonality (calculated as coefficient
of variation for annual range of temperature and precipitation) were
also adopted for reflecting the seasonal variations of these two climatic
factors. The spatial resolution of these climatic factors is 1 km.
Five important topographic variables, including elevation, slope,
aspect, vector ruggedness measure, stream power index, were collected
from the Multi-Error-Removed Improved Terrain (MERIT) DEM data
product (
) and Geomorpho90m global dataset
(
). The spatial resolution of these topographic fac-
tors is 90 m.
The variables representing vegetation growth were obtained by
using vegetation indices derived from the satellite sensors of the MODIS.
The MODIS Vegetation Indices product (MOD13A1 v061) (
)
was adopted for extracting annual mean Enhanced Vegetation Index
(EVI), annual minimum EVI, annual maximum EVI values at each
sample location. The spatial resolution of the vegetation index value for
each pixel is 500 m.
The bedrock property plays a key role in many processes at the Earth
surface, and it is also an important influencing factor for soil predictions.
We collected the rock property information from a global lithological
map database (GLiM) with a resolution of 0.5 degree (
), and extracted the value of basic lithological class at
each soil sample location.
The integration of above datasets allows for the establishment of
relationships between SOC and environmental covariates at a global
scale, providing a solid foundation for the development and training of
our deep learning model. It is noted that the set of covariates collected
above for modelling was not designed to be too large for exhaustively
covering all surface information, such as more different combinations of
features across the spectra. The reason to reduce this complexity is that
the goal of our modelling at the global scale is to generate a pre-trained
large model reflecting the basic soil
–
environment relationships. Our
emphasis is on constructing a foundational global-to-regional soil
Fig. 1.
Spatial distributions of global and regional soil sample data. a, A total number of 106,167 sampling sites (soil profiles) including soil organic carbon (SOC)
observations in topsoil were collected from WoSIS database. b-e, Four regional-scale study areas with SOC sample data that are not included in WoSIS database. These
study areas are located in Anhui province, China (region 1,
n
=
819) (b), Switzerland (region 2,
n
=
150) (c), Skjern river catchment, Denmark (region 3,
n
=
317)
(d), and New South Wales, Australia (region 4,
n
=
490) (e).
L. Zhang et al.
Geoderma
461
(2025)
117466
3
mapping approach that is transferable across regions and built on
covariates that are widely recognized as core drivers of soil formation,
broadly available and minimally uncertain. In this sense, we intention-
ally designed the scope of input variable set to include the most
fundamental and widely accessible environmental covariates as
described above. This way can make it easier for any modelers or users
to first collect the same set of basic covariate data to re-utilize the pre-
trained model we presented here, and then to adjust or extending the
model by using their local-specific datasets.
2.2. Regional-scale study areas and datasets
To evaluate the transferability of the global-scale pre-trained model,
four distinct regional-scale study areas were selected in consideration of
their varying environmental backgrounds and SOC gradients. Three of
these study areas are located in the Northern Hemisphere, and one is in
the Southern Hemisphere (
). These regions were chosen for
representing different climates, biomes and land covers, thereby
providing a robust test for applying the proposed modelling strategy
across different regional scales.
Region 1 represents the coverage of Anhui province in central-
eastern China, characterized by a subtropical humid monsoon climate
(
). The elevation is roughly from 0 to 1806 m with flat
plains in the north, low hills in the middle and rugged mountainous
terrain in the south. The average annual temperature in Anhui is around
15
◦
C, with annual precipitation ranging from 750 to 2000 mm. The land
cover types mainly consist of croplands, forests and grass and shrub-
lands. Our research group has collected 819 soil samples with SOC ob-
servations in this area (
b
) (
). Region 2 covers
the entire country of Switzerland. The climate in this region is generally
temperate but varies greatly across localities, ranging from the near-
Mediterranean climate at the southern tip to the glacial conditions on
the mountaintops. More than half of this region is dominated by the
Swiss Alps with rugged peaks, deep valleys and numerous glaciers. The
remaining areas are broadly categorized into the Swiss plateau and the
Jura mountains. The two primary terrestrial ecoregions in this region are
the western European broadleaf forests and the conifer and mixed for-
ests of the Alps (
). There are 150 soil samples
collected from the EU project Land Use/Cover Area Frame Survey
(LUCAS) (
) in this region (
c
). Region 3 is the
Skjern river catchment located in Western Jutland, Denmark. The
climate in this area is temperate maritime, with a mean annual tem-
perature of approximately 8
◦
C and annual precipitation around 990
mm. The land surface elevations are from sea level at the coast to 125 m
above sea level in the eastern area (
Jensen and Illangasekare, 2011
). The
land use is mainly agriculture, followed by grasslands and forests. There
are 317 soil samples collected in this area (
) (
d
).
Notably, the soil samples from the above three study areas are not
included in the WoSIS database. Region 4 is New South Wales in eastern
Australia. The climate varies from warm temperate in the north and east
to hot arid in the far west and subalpine in the southeastern highlands.
Annual rainfall ranges from less than 200 mm to over 2,000 mm, with
average daily maximum temperatures between 12
◦
C and 30
◦
C.
Elevation goes from sea level along the coast to over 2,000 m inland. The
land is used for large-scale agriculture (crops and livestock), forestry
(native forests and plantations), urban development in major cities, and
conservation areas like national parks (
). In this region,
490 soil samples were collected, with half included in the WoSIS data-
base and half not (
e
).
The environmental covariates in these four regions were collected
and processed to align with the datasets and variables used in the global
dataset. The water bodies and urban areas in these study areas were
excluded when applying models to generate the predicted maps of SOC.
The descriptive statistics of the soil and environmental covariates in
these regions are shown in
.
Table
1
Description
of
environmental
covariates
associated
with
soil
organic
carbon
variations
used
as
inputs
to
train
the
global-scale
pre-trained
deep
learning
model.
Category
Covariate
name
Abbreviation
Spatial
resolution
Window
size
(default)
Data
source
Reference
Reflectance
bands
NBAR
band1
(620
–
670
nm)
BAND1
500
m
5
×
5
MODIS
BRDF/Albedo
products
(MCD43A4
v061)
(
)
NBAR
band2
(841
–
876
nm)
BAND2
NBAR
band3
(459
–
479
nm)
BAND3
NBAR
band4
(545
–
565
nm)
BAND4
NBAR
band5
(1230
–
1250
nm)
BAND5
NBAR
band6
(1628
–
1652
nm)
BAND6
NBAR
band7
(2105
–
2155
nm)
BAND7
Climate
Annual
mean
temperature
BIO01
1
km
3
×
3
WorldClim
(version
2)
(
)
Temperature
Seasonality
BIO04
Annual
Precipitation
BIO12
Precipitation
Seasonality
BIO15
Topography
Elevation
ELEV
90
m
7
×
7
Geomorpho90m
(
)
Slope
SLP
Aspect
ASP
Vector
ruggedness
measure
VRM
Stream
power
index
SPI
Vegetation
Annual
mean
EVI
EVI_mean
500
m
5
×
5
MODIS
Vegetation
Indices
product
(MOD13A1
v061)
(
)
Annual
minimum
EVI
EVI_min
Annual
maximum
EVI
EVI_max
Parent
material
Bedrock
type
BEDROCK
0.5
◦
1
×
1
GLiM
(
)
Note:
Moderate
Resolution
Imaging
Spectroradiometer
(MODIS);
Nadir
Bidirectional
Reflectance
Distribution
Function
(BRDF)-Adjusted
Reflectance
(NBAR);
Enhanced
Vegetation
Index
(EVI);
Global
Lithological
Map
Database
(GLiM).
L. Zhang et al.
Geoderma
461
(2025)
117466
4
2.3. Architecture of ‘GSoilCPM
’
deep learning model
The proposed GSoilCPM model is built on a deep learning framework
to capture the complex relationships between SOC and environmental
covariates. The model uses a deep neural network architecture, specif-
ically a convolutional neural network (CNN), to extract important fea-
tures from the input data, connect and fuse these features, and finally
generate the predicted values. The overall architecture of the model is
shown in
.
The model inputs are separated by the category of covariate, with
each category having a pre-defined 2-D (dimensional) input size. This
input consists of a set of pixels in square shape with a center pixel where
a soil sample located in, and then includes pixels within a window
expanding outward from the center point, with a specific spatial reso-
lution for matching the original imagery dataset. This allows the model
to process covariates separately by considering their varying spatial
contexts and resolutions. For example, the climatic variables are often
not necessary to have a high resolution like topographic variables, as the
magnitude of variations in spatial neighborhood information is rela-
tively smaller in climatic variables than that in topographic variables.
Therefore, different covariates need be processed with their specific
window sizes and resolutions before being used as model inputs.
Each category of covariates includes multiple 2-D input variables
with the same window size and resolution, allowing them to be com-
bined into a 3-D input for that category. The CNNs were adopted to
process these inputs separately and extract their latent features. This
includes multiple times of convolutional and pooling operations, which
are effective in feature extraction (
). More details on
the operations in CNNs refer to Appendix A1. Each category of cova-
riates can be processed to produce a vector of extracted feature. All
feature vectors are then concatenated into a single, longer vector. A
second round of CNN processes this combined vector to extract the
Fig. 2.
Architecture of the Global Soil Carbon Pre-trained Model (GSoilCPM). This global-scale model for predicting soil organic carbon (SOC) is designed by using a
deep learning framework with a structure that allows for differentiated processing environmental covariates influencing SOC based on soil formation theory. The
model uses convolutional networks to extract the latent features from multiple 2-D covariate data for each input category separately, with adjustable window size to
account for spatial context and resolution of inputs. These extracted features are concatenated together and processed through convolution and fully connected layers
to extract their interactions, and then output the final predicted SOC (
̂
SOC
). The global soil profile observations and remote sensing derived environmental covariates
are collected as the global-scale database to support the training procedure of the model. The lighted arrow lines in orange represent that using backpropagation to
optimize model parameters by minimizing the loss function generated from comparing the difference between predicted and observed SOC values. (For interpretation
of the references to color in this figure legend, the reader is referred to the web version of this article.)
L. Zhang et al.
Geoderma
461
(2025)
117466
5
interactions among these features. The dropout layers are included to
prevent overfitting. We call this the feature fusion process, which sup-
ports the subsequent stage that uses a multilayer neural network (Ap-
pendix A2) for continued forward computation. Finally, fully connected
layers perform the regression to output the predicted SOC (
̂
y
).
2.4. Global-to-regional training strategy
Different from the usual way of regional soil predictions and soil
mapping based on data only in those regions, in this study, the regional
soil predictive mapping task is performed by using the regional data
with the support of global data. This modelling idea can be called as a
‘global-to-regional
’
model training strategy. The basic concept to
implement it is using domain adaptation via transfer learning. It aims to
adapt a deep learning model trained on a source domain (i.e., global-
scale soil prediction) to perform well on a target domain (i.e.,
regional-scale soil prediction), when data distribution in the two do-
mains are different but the prediction tasks are similar. Based on the
deep learning architecture, domain adaptation can be done by reusing
the learned feature representations from a data-rich source domain and
fine-tuning the model using relatively limited data from target domain
(
Ganin et al., 2016; Farahani et al., 2021
). The specific methodology and
implementation of the above basic concept are described as follows.
Typically, DSM relies on datasets in a certain study area. A data-
driven approach is often adopted, employing a machine learning
model to establish a soil
–
environment relationship (
f
) based on soil
samples and environmental covariate datasets only from that region
(
a
). This conventional approach can generate a predictive model
(
f
θ
r
:
X
r
↦
y
r
) for a region as follows:
̂
y
r
=
f
θ
r
(
X
r
)
(1)
where
X
r
represents the environmental covariates in the region;
̂
y
r
represents the predicted SOC values in the region;
θ
r
represents the
model parameters (e.g., the weights that are parameterized in a deep
neural network), which can be trained by minimizing the loss function
as follows:
̂
θ
r
=
argmin
θ
r
L
(
y
r
,
f
θ
r
(
X
r
)
)
=
argmin
θ
r
1
n
r
∑
n
r
i
=
1
(
y
(
i
)
r
−
f
θ
r
(
x
(
i
)
r
) )
2
(2)
where
L
( • )
is the loss function measuring the mean squared difference
between the predicted values
f
θ
r
(
X
r
)
and observed SOC values (
y
r
);
n
r
is
the number of samples in the region;
x
(
i
)
r
and
y
(
i
)
r
represent the envi-
ronmental covariates and the observed SOC values for the
i
-th sample in
the regional data, respectively. Here,
θ
r
is optimized from random
starting values, without any prior support from the data out of the re-
gion. The backpropagation procedure is adopted to compute the
gradient of the loss function with respect to the model parameters to
train the model (Appendix A3).
In this study, the global-to-regional training strategy is shown in
b
. We first fit a global-level soil
–
environmental relationship
(
f
θ
g
:
X
g
↦
y
g
), which is trained by minimizing the loss function on the
global dataset:
̂
θ
g
=
argmin
θ
g
L
(
y
g
,
f
θ
g
(
X
g
) )
=
argmin
θ
g
1
n
g
∑
n
g
j
=
1
(
y
(
j
)
g
−
f
θ
g
(
x
(
j
)
g
) )
2
(3)
where
X
g
and
y
g
represent the environmental covariates and the
observed SOC values for all global samples, respectively;
n
g
is the
number of global samples (generally
n
g
≫
n
r
);
x
(
j
)
g
and
y
(
j
)
g
represent the
environmental covariates and the observed SOC values for the
j
-th
sample in the global data, respectively.
Then, the global model can be fine-tuned using the dataset in a target
regional area to obtain the regional-level soil
–
environment relationship
(
f
θ
g
→
r
). Therefore, to transfer the global model to a regional setting, we
define an adaptation procedure
F
adapt
that transforms the pre-trained
global model into a regional model by considering limited regional
sample data:
f
θ
g
→
θ
r
=
F
adapt
(
f
θ
g
;
X
r
,
y
r
)
(4)
Alternatively, the adaptation can be expressed as optimizing the
regional-specific parameter shift
Δ
θ
:
θ
r
=
θ
g
+
Δ
θ
(5)
̂
Δ
θ
=
argmin
θ
g
L
(
y
r
,
f
θ
g
+
Δ
θ
(
X
r
)
)
(6)
This formalism explicitly reflects that the regional model can be built
upon the global model and optimized further with regional data, via a
training procedure of adaptation. It is noted that this training strategy
allows the regional model parameters
θ
r
to be initialized and further
optimized from the pre-trained global parameters. This training strategy
is particularly valuable for DSM, where regional datasets are often
sparse, but global datasets are more comprehensive. By employing
domain adaptation via transfer learning, it enables knowledge reuse of
the generalized soil
–
environment relationships learned globally,
thereby reducing local data requirements and enhancing predictive
Fig. 3.
The difference of basic concepts between the conventional and the proposed new soil mapping methods. (a) The conventional data-driven approach which
trains a soil
–
environmental relationship (
f
r
) at the regional level only using the soil sample and environmental covariate datasets in a certain region. (b) The proposed
new soil prediction/mapping approach first fits a global-level soil
–
environment relationship (
f
g
) (it can be fitted by the proposed GSoilCPM that is illustrated in
), and then fine-tunes the model by using the dataset in a target regional area to obtain the regional-level soil
–
environment relationship (
f
g
→
r
). An adaptation
procedure
F
adapt
is used to transform the pre-trained global model into a regional model by considering limited regional sample data (
Data
r
).
L. Zhang et al.
Geoderma
461
(2025)
117466
6
performance.
The previously described global soil profile data with SOC observa-
tions and remote sensing derived environmental covariates, serve as the
global- and regional scale database to support this training process. The
backpropagation is used to optimize model parameters based on the loss
function, which is represented as the lighted arrow lines in orange color
in
2.5. Evaluation of SOC predictions
A ten-fold cross-validation was adopted to assess the prediction ac-
curacy by using different modelling methods. This approach helps pre-
vent bias in evaluation caused by the possible overfitting of the model to
a single validation set. In each region, all samples were partitioned into
ten subsets. Nine subsets were used as the training data to fit the model,
and the model performance was validated on the remaining fold. The
final assessment of the model was based on the mean accuracy across the
ten validation sets. To ensure a representative spatial distribution of the
training and validating sample sets, we adopted a global grid with 5-by-
5-degree latitude-longitude space, covering the globe, and then split the
training and validating sample sets in each grid cell for ensuring uniform
spatial coverage. This procedure was repeated ten times, and each time
using a different subset for validation. The accuracy metrics of the mean
absolute error (MAE) and the concordance correlation coefficient (CCC)
were computed on the validation samples to assess the model perfor-
mance:
MAE
=
∑
n
i
=
1
|
y
i
−
̂
y
i
|
n
(7)
CCC
=
2
r
σ
y
σ
ˆ
y
σ
2
y
+
σ
2
ˆ
y
+ (
y
−
̂
y
)
2
(8)
where
n
is the number of validation samples;
y
i
and
̂
y
i
are
i
-th observed
and predicted SOC value respectively;
y
and
̂
y
are the averages of the
observed and predicted values;
σ
y
and
σ
ˆ
y
are the corresponding standard
deviations; and
r
is the correlation coefficient value between the pre-
dicted and observed values. A smaller value of MAE or a larger value of
CCC means a higher prediction accuracy.
The random forest (RF) model was taken as a reference model for
comparing with our proposed modelling approach. As an ensemble
model, RF generates random subsets of data to train different base
learners and then aggregates their predictions, which allows it to have a
good ability in generalization. Considering RF has been tested to be
superior to many other conventional machine learning models for soil
mapping tasks (
Brungard et al., 2015; Heung et al., 2016; Zhang et al.,
), it is reasonable and more challenging to adopt it as an reference
model or a baseline to evaluate the performance of the proposed deep
learning model. The RF model was trained only on regional data,
considering it exemplifies a conventional
“
local-only
”
machine learning-
based modelling strategy, which is commonly used in many regional
DSM tasks. To illustrate the efficiency and advantages of using pre-
trained global-scale model for enhancing regional-scale SOC pre-
dictions, the deep learning model with the same architecture and built
on regional data but without using the pre-training was also taken for
comparing. The data split for training and validation in the cross-
validation was identical for both the RF and DL models, ensuring a
consistent basis for model evaluation and comparison. We further
compared our regional modelling results with the SoilGrids product to
highlight how widely used global soil data products suffer from potential
biases in regional-scale predictions.
To test how modelling improvement (measured as the percentage
improvement of prediction accuracy from GSoilCPM versus the baseline
RF model) changes in response to different regional sample sizes and
baseline accuracies, we determined different numbers of sample data for
training models and thus to detect the potential relationships. We
randomly selected different proportions of sample data (ranging from
10% to 100% with an interval of 10%) to the total available training data
in each region. For each sample size, we repeated the random sampling
process 100 times, allowing us to observe statistical differences in the
modelling results across sample sizes. We further aggregated all
modelling results to analyze the relationship between model improve-
ments and changes in the baseline accuracy of SOC prediction in each
study area.
3. Results and Discussion
3.1. Assessment of global-scale modelling
The global-scale deep learning model (GSoilCPM) exhibited
competitive model performance of SOC compared with the widely used
RF model at the global scale (
), both of which were trained on the
same WoSIS global soil sample data and environmental covariates.
Specifically, cross-validation results illustrate that the deep learning
model achieved an accuracy of 7.15 g kg
−
1
in mean absolute error
(MAE) and 0.57 in concordance correlation coefficient (CCC), while the
RF model had a MAE of 7.56 g kg
−
1
in and a CCC of 0.57. The RF model
is well-known for its strong generalization capability and has been
widely used. While previous studies have showed challenges in sur-
passing RF performance with deep learning models (
), our results indicate that deep learning model can
achieve a comparable or slightly better prediction accuracy in both
metrics when large amounts of global-scale data are available. Both
models effectively captured the global variability in SOC by modelling
the complex relationships between SOC and various environmental
covariates.
3.2. Global-scale pre-trained GSoilCPM allows enhancement in regional-
scale SOC predictions
SOC prediction accuracies for different models across the four re-
gions were assessed using ten-fold cross-validation (
). For com-
parison, we also evaluated the accuracy of the SoilGrids product,
validated by our regional sample data (column 1 in
). In general,
across all four study areas, models trained on regional sample data
produced significantly better predictions than using SoilGrids directly.
This is because SoilGrids was derived from a model trained on a global
soil sample database (i.e., WoSIS data) that does not include samples
from our study areas. While previous studies have shown that SoilGrids
can provide good SOC predictions (
Hengl et al., 2017; Poggio et al.,
), our findings suggest that its global model cannot guarantee
acceptable accuracy in regions lacking WoSIS sample data.
Comparing the baseline RF model with the DL model without pre-
training (i.e., models are trained only based on the data in a region)
(columns 2 and 3 in
), we found that the proposed DL model ar-
chitecture (without pre-training) achieved competitive result compared
to RF, even with limited regional sample data. More importantly, we
found that when we used the DL model with pre-training, which means a
pre-trained GSoilCPM was constructed and then it was transferred into
these regional study areas, model performance shows a large improve-
ment (column 4 in
). These results support our hypothesis that the
regional-scale soil carbon predictions can be enhanced by transferring
global-scale soil
–
environment relationships via using a DL-based
modelling approach. Specifically, in four case studies from Region 1 to
Region 4, the metrics of MAE (unit: g kg
−
1
) decreased by 0.13, 0.14, 0.08
and 0.04, and the metrics of CCC increased by 0.43, 2.23, 1.75 and 0.40
when comparing the RF and fine-tuned pre-trained DL models, respec-
tively. Similar improvements in accuracy metrics were observed in all
other study areas. In general, the percentage improvements are 10.21%,
12.92%, 11.25% and 9.32% in MAE and 41.94%, 32.56%, 36.36% and
5.33% in CCC for Regions 1
–
4, respectively. The average improvement
across all regions is 10.93% (decreased by 1.20 g kg
−
1
) in MAE and
L. Zhang et al.
Geoderma
461
(2025)
117466
7
29.04% (increased by 0.10) in CCC. These results substantiate that, by
leveraging the GSoilCPM deep learning model pre-trained based on a
vast repository of global SOC data, the model can effectively transfer its
nuanced understanding of complex global soil
–
environment relation-
ships to capture regional soil variations.
When comparing mapping results from SoilGrids, RF models, and
pre-trained DL models, similar general pattern of SOC variation across
space are observed in each region. However, substantial differences in
value ranges and some local details of SOC distributions are observed
(
). It shows that SoilGrids tend to overestimate SOC in regions with
lower SOC (relative to the global mean), and underestimate SOC in re-
gions with higher SOC. For example, across the mountainous areas of
Switzerland (Region 2) and arid areas in New South Wales, Australia
(Region 4), SoilGrids shows a large bias compared to the prediction
maps generated using local sample data (
d-f, j-l
). This bias may
result from the RF model
’
s tendency to compress the range of predicted
values due to its intrinsic method of averaging multiple predictions from
individual trees, smoothing out extremes and reducing variability in the
predictions (
). This problem is more prominent
when comparing the regional and global mapping results, as the
regionally extreme low and high SOC values might be compressed to
reduce the global-level bias in average.
The mapping results from RF and GSoilCPM (pre-trained DL model)
also show notable differences. The SOC maps derived from the pre-
trained DL model show more detailed spatial heterogeneity, especially
can be found in mountainous and rugged areas across the southern part
in Region 1 and 2. This can be partly explained by the fact that the
proposed DL model can produce a higher spatial resolution of mapping
results. This advantage comes from the GSoilCPM model
’
s ability to
incorporate covariate data with different spatial contexts and different
original resolutions, whereas traditional models usually require all
covariates to be pre-processed to the same resolution. This pre-
processing step may lead to the loss of detailed information from local
neighborhood pixels (
Wadoux, 2019; Yang et al., 2021a; Zhang et al.,
), such as high resolution topographic variables. Due to that
different factors may influence SOC at different spatial scales (
et al., 2010; Tan et al., 2024
), resampling all the covariates into a same
fixed resolution may cause mismatches between the affecting scale of
environmental factors on SOC. Moreover, the density distribution of
SOC values mapped by the pre-trained DL model better matches the
sample-level SOC distribution than that from RF model (
), and
the DL model also demonstrated higher validation accuracy. Therefore,
the prediction maps of SOC generated by the pre-trained DL models can
be considered more reliable. By combining global and regional soil and
environmental datasets via transferring a pre-trained DL model, the final
prediction maps reconciled the global knowledge of soil
–
environment
relationships extracted from global databases with the region-specific
SOC patterns derived from local observations.
3.3. Modelling improvement related with sample size and baseline
accuracy
The relationships between model improvement (i.e., the accuracy
improvement of the pre-trained DL model over the baseline as repre-
sented by the RF model), regional sample size, and baseline accuracy are
crucial for understanding how the performance of GSoilCPM changes
under different conditions. In our analysis, we observed that applying
GSoilCPM in all four regions consistently resulted in better performance
on average, compared to the baseline (
a
). More importantly, it
shows that the degree of enhancement in SOC prediction particularly
became greater as the number of samples in a region decreased (
).
This trend highlights the advantages of transferability and robustness of
the GSoilCPM, particularly when regional sample data are limited. For
instance, in Region 2 and 3, models trained with less than half the total
sample size reported an over 50% model improvement on average. This
trend was less evident in Region 1, where the degree of improvement
remained relatively consistent across different sample sizes. This is
mainly because Region 1 has the largest number of sample data
compared to the other three regions, leading to less contrast in model
performance when using samples sizes greater than ten percentage of
the total. In addition, it also indicates that a larger number of available
regional samples may sometimes continuously contribute to further
improvement in DL model fitting.
The observed improvements are also closely linked to baseline ac-
curacy
—
specifically, the prediction accuracy derived from the RF
model. In cases where baseline accuracy was lower, the degree of pre-
diction enhancement when utilizing GSoilCPM tended to be higher
(
). Conversely, in areas with a higher baseline accuracy, the
relative improvements were more modest. This phenomenon was found
in all four study areas. Interestingly, the slope of this negative rela-
tionship was more pronounced when the mean prediction accuracy of
the baseline RF model was low, such as the steep slope shown in Region
2 and 3 when the CCC of RF model result was below 0.3 (
).
Generally, these results demonstrate that as the number of soil samples
decreases, incorporating a pre-trained DL model, such as GSoilCPM
proposed in this study, not only enhances SOC prediction accuracy but
does so more effectively where the model performance is relatively low
when only using regional sample data.
3.4. Modelling from globe to regions: A new paradigm for soil mapping
The modelling strategy employed in this study represents a potential
Fig. 4.
The validation accuracies of modelling soil organic carbon content (SOC) by using random forest model (a) and the proposed GSoilCPM deep learning model
(b) at the global scale. All dots are the observed and predicted SOC values for all validation samples from ten-fold cross-validation.
L. Zhang et al.
Geoderma
461
(2025)
117466
8
paradigm shift in soil mapping, wherein global-scale pre-training a DL
model can serve as a domain adaptation model for more accurate
regional-scale soil predictions. By generating the interconnectedness of
SOC and environmental covariates on a global scale, we can establish
frameworks that facilitate the regional application of soil mapping via
the GSoilCPM architecture. Traditionally, regional soil mapping ap-
proaches often rely solely on limited regional datasets, which may
introduce biases and limit the generalizability of models when regional
sample data are insufficient. Conversely, the modelling approach based
on GSoilCPM leverages soil
–
environment relationships obtained from
the vast amount of globally available soil and environmental data,
providing a helpful foundation for regional fine-tuning.
The proposed ‘global-to-regional
’
modeling strategy derives its
effectiveness from the theoretical plausibility of the following basic
concepts. First, soil variations are fundamentally shaped by long-term
interactions with climatic, topographic, and ecological factors that
exhibit consistent patterns across scales. These broad-scale
soil
–
environment relationships, often governed by macro geographical
and ecological regularities, can be learned from globally distributed data
and serve as transferable knowledge. Second, deep learning models are
well-suited for capturing such abstract and hierarchical patterns,
enabling them to inductively generalize from diverse environmental
contexts. Thus, the integration of transferable global knowledge with
region-specific fine-tuning offers a new solution for enhancing regional
DSM.
Furthermore, the search process for the optimal hypothesis space (i.
e., a set of all possible functions or models that the learning algorithm
can produce, and each function in this space represents a potential so-
lution to the problem, such as the problem of SOC prediction) regarding
soil
–
environment relationships can be markedly improved when using
GSoilCPM.
conceptually illustrates why GSoilCPM offers an
effective modelling strategy for regional soil predictions. In traditional
regional modelling, machine learning algorithms search for an optimal
predictive function within a hypothesis space using only local data. This
Fig. 5.
Comparisons of validation accuracies for predicting soil organic carbon content (SOC) using different modelling strategies across four regional-scale study
areas. Each row represents the results for one regional study area. The cross-validation accuracies for SoilGrids (a, e, i, m), random forest (RF) models (b, f, j, n), deep
learning (DL) models without pre-training (c, g, k, o), and pre-trained DL model (i.e., GSoilCPM) (d, h, l, p) are shown in the first, second, third and fourth columns,
respectively.
L. Zhang et al.
Geoderma
461
(2025)
117466
9
process is prone to overfitting or convergence to suboptimal local
minima, particularly when observations are sparse. By contrast, the
GSoilCPM framework leverages a globally pre-trained model that en-
codes generalized soil
–
environment relationships. This global knowl-
edge effectively constrains the hypothesis space, guiding the regional
model initialization closer to the optimal solution. As a result, the sub-
sequent fine-tuning process on regional data becomes more stable,
sample-efficient, and less susceptible to possible poor convergence. This
global-to-regional transfer learning approach provides a plausible way
to reduce optimization bias in regional-scale soil mapping.
The advantage in adopting this modelling strategy also stems from
the fact that the environmental conditions in a certain region often
overlap with those collected from the global database. Using principal
component analysis (PCA) to transform the soil and environmental
Fig. 6.
Mapping results of predicted soil organic carbon content (SOC) derived from SoilGrids (a, d, g, j), random forest (RF) model (b, e, h, k), and the proposed pre-
trained deep learning (DL) model (c, f, i, l). Each row shows the comparison of SOC maps in a regional area. The spatial resolution of RF and DL model derived maps
is 90 m, compared to the 250 m resolution for SoilGrids product.
L. Zhang et al.
Geoderma
461
(2025)
117466
10
covariate data into a two-dimensional feature space (
) shows that
all four regions overlap with the global data, although the extent and
location of these overlaps vary. This suggests that some of the knowl-
edge or relationships generated from global data can complement
regional-scale SOC modelling, thereby reducing the difficulty in estab-
lishing soil
–
environment relationships given limited samples at regional
scales. As an illustration shown in
, by initiating the search from a
hypothesis space generated by a pre-trained model, optimization algo-
rithms can converge on the best-fitting solution more rapidly and
effectively. This pre-defined space incorporates knowledge accumulated
from global datasets, providing a robust foundation for exploring model
parameters in a local area. It can help the model optimization start from
a pre-trained global model, which narrows down the initial hypothesis
space and makes it easier to find a good fit for the regional data. In
contrast, searches originating from a random starting point in the uni-
versal space often yield longer search times and may struggle to reach
the optimal result due to the vast number of potential configurations
involved. A very large initial hypothesis space might lead to higher
possibility of overfitting, where the model may learn the noise in the
limited training data rather than the underlying patterns.
It is also crucial to recognize the challenges associated with directly
relying on global soil map products such as SoilGrids for regional-scale
studies. Our results indicate that the predicted global soil map products
may exhibit considerable biases in regions where local sample data were
not included in the model training process, despite it has been validated
to have strong global-level model performance. Promisingly, GSoilCPM
is designed to overcome these biases by effectively integrating local data
alongside global observations.
3.5. Limitations and future horizons
The current implementation of GSoilCPM primarily focuses on
Fig. 7.
Accuracy improvement with varying sample sizes (a) and baseline accuracies (b). The accuracy improvement is calculated as the percentage increase by using
the pre-trained deep learning model (GSoilCPM) compared to random forest (RF) model. The accuracy (using the metrics of CCC here) of RF model is adopted as the
baseline accuracy. Error bars show 90% percentile intervals of accuracy improvement for all modelling runs for each sample size.
Fig. 8.
Why GSoilCPM offers an effective modelling strategy for regional soil predictions. The background color gradient represents the variation of model accuracy
in the hypothesis space (a searchable space of the model function). The optimal location indicates the best model function. Building a pre-trained global model to
capture the general soil
–
environment relationships can help narrow the scope of searching regional-scale predictive function. The global-to-regional modelling
strategy can reduce the risk of falling into local minima when fitting a machine learning (ML) model based on region data alone.
L. Zhang et al.
Geoderma
461
(2025)
117466
11
predicting a single soil property, specifically SOC in the topsoil, which is
widely concerned for soil health and carbon sequestration. While this
model has not been applied to predict other soil properties, our bench-
mark test suggests a large potential of adopting this modelling concept
for enhancing regional predictions of all soil properties. This prompts
further investigation to ascertain whether similar modelling techniques
can generalize effectively to other soil mapping tasks. Additionally, how
to modify the model to take account for the variations of soil properties
across different soil depths is an open question. Developing a whole-
profile pre-trained model certainly needs further explored in future
work. Furthermore, the next version of the GSoilCPM framework could
be expanded by adding a modular sub-network designed to handle
region-specific inputs (e.g., high-resolution agricultural management
data), which can be integrated during the regional fine-tuning stage
when reliable local data are available.
While the statistical average of model accuracy show improvement
across all four study areas, the results reveal a high degree of variability.
The large variance in the degree of model improvement suggest that
enhancements cannot be consistently attained across all modelling in-
stances. Some regions or data availability conditions could exhibit less
pronounced gains. This variability necessitates caution when trans-
ferring models and suggests that further studies are needed to identify
situations where the pre-trained model may underperform.
There are promising horizons for future research and development
related to the GSoilCPM model. One key direction involves expanding
the model to predict multiple soil properties simultaneously. Imple-
menting a multi-task learning framework could allow us to generalize its
findings across various soil attributes and depths. Moreover, ongoing
investigations into optimizing the model performance under different
regional conditions should continue. This includes determining the
optimal sizes and resolutions of model inputs, and expanding the model
structure to allow the inclusion of specific local dataset, such as the extra
inputs reflecting human activities for a region. Maximizing the model
applicability from these perspectives will be a promising development
direction to improve GSoilCPM
’
s predictability across diverse contexts.
4. Conclusion
This study presents GSoilCPM, an innovative deep learning frame-
work designed to advance the prediction and mapping accuracy of soil
organic carbon at regional scales. By integrating global soil datasets with
environmental covariates derived from satellite-based observations, our
proposed modelling strategy effectively addresses a critical knowledge
gap on how to use the global-scale soil
–
environment relationships
learned from global databases to enhance the regional-scale soil pre-
dictions with limited regional soil data. The results in four study areas
prove that GSoilCPM, when fine-tuned for regional contexts, signifi-
cantly outperforms models trained exclusively on regional data. The
relationships between model improvement, regional sample size, and
baseline accuracy that without using pre-trained model, revealed in our
study, are crucial for understanding how the performance of GSoilCPM
changes under different conditions. Our analyses also indicate that, re-
gions with fewer soil samples and/or lower baseline accuracy can
benefit more from using the pre-trained global-scale model, under-
scoring the importance of leveraging global-scale observations to over-
come local data limitations.
In summary, the insights gained from this study are valuable for
better informing the next generation of soil modelling and mapping
methodology. The advent of the modelling approach using GSoilCPM
signifies a transformative advancement in soil mapping, particularly in
enhancing regional soil predictions by bridging global- and regional-
scale soil
–
environment relationships. The breakthroughs offered by
the ‘global-to-regional
’
transfer deep learning model extend beyond
mere predictive enhancements, it represents a potential new paradigm
in the methodology of soil mapping itself. By embracing a model that
synergizes local insights with global knowledge, scientists and land use
managers can achieve a data- and model-reusable approach to under-
stand soil variations more accurate. Such integration not only optimizes
existing information but also drastically improves the modelling effi-
ciency in regional soil predictions, positioning GSoilCPM as a new
pivotal tool in advancing the field of soil monitoring and land
management.
CRediT authorship contribution statement
Lei Zhang:
Writing
–
review
&
editing, Writing
–
original draft,
Visualization, Validation, Supervision, Software, Resources, Project
administration, Methodology, Investigation, Funding acquisition,
Formal analysis, Data curation, Conceptualization.
Lin Yang:
Writing
–
review
&
editing, Supervision, Resources, Project administration,
Investigation, Funding acquisition, Conceptualization.
Yuxin Ma:
Writing
–
review
&
editing, Validation, Resources, Investigation, Data
curation.
A-Xing Zhu:
Writing
–
review
&
editing, Resources.
Ren Wei:
Validation, Software, Data curation.
Jie Liu:
Validation, Software.
Mogens H. Greve:
Writing
–
review
&
editing, Resources, Investigation,
Data curation.
Chenghu Zhou:
Writing
–
review
&
editing, Validation,
Supervision, Resources, Project administration, Investigation, Funding
acquisition, Conceptualization.
Fig. 9.
An illustration explaining the model optimization when taking advantages of GSoilCPM to search the optimal hypothesis (
H
) of regional soil
–
environment
relationships.
L. Zhang et al.
Geoderma
461
(2025)
117466
12
Declaration of competing interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Acknowledgements
This work was supported by the National Key Research and
Development Program Plan (Grant No. 2022YFC3800802), the National
Natural Science Foundation of China (Grant No. 42471468), the Leading
Funds for the First-Class Universities (020914912203 and
020914902302). Jie Liu was supported by the Postgraduate Research
&
Practice Innovation Program of Jiangsu Province (Grant No.
KYCX25_0209). Lei Zhang acknowledges the support from the U.S.
Department of Energy, Office of Science, Office of Biological and Envi-
ronmental Research, under Award Number DE-AC02-05CH11231.
Appendix A. . More details of GSoilCPM
Appendix A. 1. Convolutional neural networks in GSoilCPM
The convolutional neural networks (CNNs) are adopted in the proposed GSoilCPM framework for extracting latent features from spatial infor-
mation of environmental covariates. CNNs are designed to handle grid-like data such as raster images. The convolutional layer is the most important
building block of a CNN. The convolution operation applies a filter (kernel) over the input. For example, if there is a 2-D data used as the input, a 2-D
kernel can be adopted to detect features of the input data (
). The kernel is a set of weights, which allows to detect features in the input data.
Each unit in the output is a weighted sum of a local patch of input values, defined as:
z
i
,
j
= (
x
*
k
)
i
,
j
=
∑
h
u
=
1
∑
w
v
=
1
x
i
ʹ
,
j
ʹ
•
w
u
,
v
with
{
i
ʹ
=
u
•
s
h
+
h
−
1
j
ʹ
=
v
•
s
w
+
w
−
1
(1)
where
z
i
,
j
is the output of the unit located in row
i
, column
j
in feature map of a convolutional layer;
h
and
w
are the height and width of the kernel
(receptive field);
s
h
and
s
w
are the vertical and horizontal strides;
x
i
ʹ
,
j
ʹ
is the unit located at row
i
ʹ
, column
j
ʹ
in the input;
w
u
,
v
is the weight between any
unit in feature map and its input located at row
u
, column
v
(relative to the unit
’
s receptive field). The local weighted sum is then passed through a non-
linear transfer function. The outputted units in a convolutional layer are organized and are referred to as the feature map. All units in a feature map
share the same kernel. This shared-weights strategy was used because we consider the local statistics of input images and their signals are invariant to
location (
Goodfellow et al., 2016; LeCun et al., 2015
). Most importantly it means once the CNN has learned to recognize a pattern in one location, it
can recognize it in any other location. In practice, multiple kernels are used in each convolutional layer to extract different types of spatial features,
making it capable of detecting multiple features simultaneously in the input data.
After the convolution operation, the max pooling operation is a useful further stage. The role of the pooling layer is to merge semantically similar
features into one. This operation can reduce the computational load, the memory usage, and the number of parameters (thereby limiting the risk of
overfitting). Max pooling slides a window over the feature map and aggregate values into the maximum value from each window (
).
In a typical CNN architecture, two or three stages of convolution, non-linearity and pooling are stacked, then followed by more convolutional
layers, non-linearity and pooling. In our study, CNNs allow GSoilCPM to learn spatial features from different environmental covariates that influence
soil, such as terrain structure, vegetation patchiness, and climatic gradients, directly from raster inputs with different spatial resolutions.
Fig. A1.
The convolution layer (a) and the pooling layer (b) as the core building blocks in a convolutional neural network (CNN).
Appendix A. 2. Multi-layer neural network and forward computation
As all feature vectors extracted by CNNs are then concatenated into a single and longer vector, this allows the multi-layer neural network to process
this vector for extracting the interactions among these features. A perceptron is a fundamental building block in multi-layer neural network (
a
).
It is essentially a single-layer neural network that can compute a weighted sum of its inputs, and then uses an activation function to produce an output:
h
=
ϕ
(
w
T
•
x
+
b
)
(2)
where
x
is the input vector;
w
provides the weights of a linear transformation and
b
is the bias;
ϕ
( • )
represents an activation function which is often set
to be a nonlinear function such as the rectified linear activation function (ReLU);
h
is the output.
A multi-layer neural network is composed of an input layer, one or more hidden layers, and a final output layer (
). All units except a bias in
L. Zhang et al.
Geoderma
461
(2025)
117466
13
a layer are fully connected to the previous layer. This multi-layer structure enables learning of abstract, hierarchical representations, essential for
generating the complex and nonlinear relationships among multiple environmental covariates influencing soil variations.
Fig. A2.
The perceptron (a) as a fundamental building block in a multi-layer neural network (b).
Appendix A. 3. Backpropagation and gradient-based optimization
Training the deep neural network involves minimizing the error between predicted and observed SOC values using backpropagation and gradient
descent. The prediction error can be calculated using the mean squared error which is considered as the loss function
L
(described in
). The
backpropagation procedure to compute the gradient of
L
with respect to the weights in GSoilCPM is fundamentally an application of the chain rule for
derivatives. The key insight is that the gradient of the loss with respect to the model parameters in a layer can be computed by working backwards from
the gradient with respect to the output of that layer. According to Taylor
’
s theorem, for a differentiable function such as
L
(
θ
)
, a small perturbation
s
leads to:
L
(
θ
+
s
) ≈
L
(
θ
) +
s
• ∇
L
(
θ
)
(3)
where
∇
L
is the gradient of
L
. This motivates the use of gradients in minimizing
L
. In neural networks, the chain rule of derivatives allows
backpropagation equation can be applied repeatedly to propagate gradients through all layers. In gradient descent, the first order gradient is used, and
we set:
s
= −
η
• ∇
L
(
θ
)
(4)
where
η
(
η
>
0) is called the
“
step size
”
or
“
learning rate
”
, which is usually set to be small to ensure
L
(
θ
+
s
) ≤
L
(
θ
)
. The direction of the negative
gradient moves the function
’
s output toward the local minimum. Therefore, the model parameters
θ
can be updated by:
θ
←
θ
−
η
•
∂
L
∂
θ
(5)
There are some faster and more stable convergence than regular gradient descent optimizer. In this study, we adopted Adam optimizer (
), which adjusts the learning rate adaptively during training.
Appendix B. Supplementary data
Supplementary data to this article can be found online at
https://doi.org/10.1016/j.geoderma.2025.117466
Data availability
All data used in the analysis are publicly accessible. The WoSIS soil
profile data were obtained from a published snapshot (September 2019)
archived at
https://doi.org/10.17027/isric-wdcsoils.20190901
. The
SoilGrids (version 2.0) database can be accessed at
.
The climate data are available in the WorldClim version 2.0 database
(
). The global lithological map database is avail-
able at
https://doi.org/10.1594/PANGAEA.788537
. The global digital
elevation model (DEM) data are available at
. The global dataset comprising of multiple
topographic features (Geomorpho90m) is obtained from
. The MODIS BRDF/Albedo and vegetation
indices products are available at
https://ladsweb.modaps.eosdis.nasa.
.
The codes for processing data have been deposited into a repository
at
https://github.com/leizhang-geo/GSoilCPM.git
. The pre-trained
model (GSoilCPM) is also stored in this repository.
L. Zhang et al.
Geoderma
461
(2025)
117466
14
References
Amatulli, G., McInerney, D., Sethi, T., Strobl, P., Domisch, S., 2020. Geomorpho90m,
empirical evaluation and accuracy assessment of global high-resolution
geomorphometric layers. Sci. Data 7, 162.
https://doi.org/10.1038/s41597-020-
Awais, M., Naseer, M., Khan, S., Anwer, R.M., Cholakkal, H., Shah, M., Yang, M.-H.,
Khan, F.S., 2023. Foundational Models Defining a New Era in Vision: A Survey and
Outlook. https://doi.org/10.48550/arXiv.2307.13721.
Batjes, N.H., Calisto, L., de Sousa, L.M., 2024. Providing quality-assessed and
standardised soil data to support global mapping and modelling (WoSIS snapshot
2023). Earth Syst. Sci. Data 16, 4735
–
https://doi.org/10.5194/essd-16-4735-
Batjes, N.H., Ribeiro, E., Van Oostrum, A., 2020. Standardised soil profile data to support
global mapping and modelling (WoSIS snapshot 2019). Earth Syst. Sci. Data 12,
299
–
320.
https://doi.org/10.5194/essd-12-299-2020
.
Batjes, N.H., Ribeiro, E., van Oostrum, A., Leenaars, J., Hengl, T., Mendes de Jesus, J.,
2017. WoSIS: providing standardised soil profile data for the world. Earth Syst. Sci.
Data 9, 1
–
14.
https://doi.org/10.5194/essd-9-1-2017
Behrens, T., Zhu, A.-X., Schmidt, K., Scholten, T., 2010. Multi-scale digital terrain
analysis and feature selection for digital soil mapping. Geoderma 155, 175
–
185.
https://doi.org/10.1016/j.geoderma.2009.07.010
.
Bishop, T.F.A., McBratney, A.B., Laslett, G.M., 1999. Modelling soil attribute depth
functions with equal-area quadratic smoothing splines. Geoderma 91, 27
–
45.
https://doi.org/10.1016/S0016-7061(99)00003-8
Bliss, N.B., Waltman, S.W., Petersen, G.W., 1995. Preparing a soil carbon inventory for
the United States using geographic information systems, in: Soils and Global Change.
CRC Press, Boca Raton, FL, pp. 275
–
295.
Brevik, E.C., Pereira, P., Mu
˜
noz-Rojas, M., Miller, B.A., Cerd
`
a, A., Parras-Alc
´
antara, L.,
Lozano-García, B., 2017. Chapter 1 - Historical Perspectives on Soil Mapping and
Process Modeling for Sustainable Land Use Management, in: Pereira, P., Brevik, E.C.,
Mu
˜
noz-Rojas, M., Miller, B.A. (Eds.), Soil Mapping and Process Modeling for
Sustainable Land Use Management. Elsevier, pp. 3
–
28. https://doi.org/10.1016/
B978-0-12-805200-6.00001-3.
Brungard, C.W., Boettinger, J.L., Duniway, M.C., Wills, S.A., Edwards, T.C., 2015.
Machine learning for predicting soil classes in three semi-arid landscapes. Geoderma
239, 68
–
83.
https://doi.org/10.1016/j.geoderma.2014.09.019
.
Chen, S., Arrouays, D., Leatitia Mulder, V., Poggio, L., Minasny, B., Roudier, P.,
Libohova, Z., Lagacherie, P., Shi, Z., Hannam, J., Meersmans, J., Richer-de-Forges, A.
C., Walter, C., 2022. Digital mapping of GlobalSoilMap soil properties at a broad
scale: a review. Geoderma 409, 115567.
.
Cui, W., Yang, L., Zhang, L., Yang, C., Zhu, C., Zhou, C., 2025. A novel approach of
generating pseudo revisited soil sample data based on environmental similarity for
space-time soil organic carbon modelling. Int. J. Appl. Earth Obs. Geoinf. 139,
104542.
https://doi.org/10.1016/j.jag.2025.104542
Dai, Y., Shangguan, W., Wei, N., Xin, Q., Yuan, H., Zhang, S., Liu, S., Lu, X., Wang, D.,
Yan, F., 2019. A review of the global soil property maps for Earth system models.
Soil 5, 137
–
158.
https://doi.org/10.5194/soil-5-137-2019
Didan, K., 2021. MODIS/Terra Vegetation Indices 16-Day L3 Global 500m SIN Grid
V061. https://doi.org/10.5067/MODIS/MOD13A1.061.
Dinerstein, E., Olson, D., Joshi, A., Vynne, C., Burgess, N.D., Wikramanayake, E.,
Hahn, N., Palminteri, S., Hedao, P., Noss, R., Hansen, M., Locke, H., Ellis, E.C.,
Jones, B., Barber, C.V., Hayes, R., Kormos, C., Martin, V., Crist, E., Sechrest, W.,
Price, L., Baillie, J.E.M., Weeden, D., Suckling, K., Davis, C., Sizer, N., Moore, R.,
Thau, D., Birch, T., Potapov, P., Turubanova, S., Tyukavina, A., de Souza, N.,
Pintea, L., Brito, J.C., Llewellyn, O.A., Miller, A.G., Patzelt, A., Ghazanfar, S.A.,
Timberlake, J., Kl
¨
oser, H., Shennan-Farp
´
on, Y., Kindt, R., Lilles
ø
, J.-P.-B., van
Breugel, P., Graudal, L., Voge, M., Al-Shammari, K.F., Saleem, M., 2017. An
Ecoregion-based Approach to Protecting half the Terrestrial Realm. Bioscience 67,
534
–
545.
https://doi.org/10.1093/biosci/bix014
.
FAO, 2012. Harmonized World Soil Database v1.2 [WWW Document]. URL http://
webarchive.iiasa.ac.at/Research/LUC/External-World-soil-database/HTML/
(accessed 6.17.22).
Farahani, A., Voghoei, S., Rasheed, K., Arabnia, H.R., 2021. A Brief Review of Domain
Adaptation. In: Stahlbock, R., Weiss, G.M., Abou-Nasr, M., Yang, C.-.-Y., Arabnia, H.
R., Deligiannidis, L. (Eds.), Advances in Data Science and Information Engineering.
Springer International Publishing, Cham, pp. 877
–
894.
.
Fick, S.E., Hijmans, R.J., 2017. WorldClim 2: new 1-km spatial resolution climate
surfaces for global land areas. Int. J. Climatol. 37, 4302
–
4315.
Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M.,
Lempitsky, V., 2016. Domain-Adversarial Training of Neural Networks. J. Mach.
.
Goodfellow, I., Courville, A., Bengio, Y., 2016. Deep learning. The MIT Press, Cambridge,
Massachusetts, Adaptive computation and machine learning
Guo, M., Yang, L., Zhang, L., Shen, F., Meadows, M.E., Zhou, C., 2025. Hydrology,
vegetation, and soil properties as key drivers of soil organic carbon in coastal
wetlands: a high-resolution study. Environ. Sci. Ecotechnol. 23, 100482.
.
Hartmann, J., Moosdorf, N., 2012. The new global lithological map database GLiM: a
representation of rock properties at the Earth surface. Geochem. Geophys. Geosyst.
13.
https://doi.org/10.1029/2012GC004370
He, X., Yang, L., Li, A., Zhang, L., Shen, F., Cai, Y., Zhou, C., 2021. Soil organic carbon
prediction using phenological parameters and remote sensing variables generated
from Sentinel-2 images. Catena 205, 105442.
.
Helfenstein, A., Mulder, V.L., Heuvelink, G.B.M., Hack-ten Broeke, M.J.D., 2024. Three-
dimensional space and time mapping reveals soil organic matter decreases across
anthropogenic landscapes in the Netherlands. Commun. Earth Environ. 5, 1
–
16.
https://doi.org/10.1038/s43247-024-01293-y
.
Hengl, T., Heuvelink, G.B.M., Stein, A., 2004. A generic framework for spatial prediction
of soil variables based on regression-kriging. Geoderma 120, 75
–
10.1016/j.geoderma.2003.08.018
Hengl, T., de Jesus, J.M., Heuvelink, G.B.M., Gonzalez, M.R., Kilibarda, M., Blagoti
´
c, A.,
Shangguan, W., Wright, M.N., Geng, X., Bauer-Marschallinger, B., Guevara, M.A.,
Vargas, R., MacMillan, R.A., Batjes, N.H., Leenaars, J.G.B., Ribeiro, E., Wheeler, I.,
Mantel, S., Kempen, B., 2017. SoilGrids250m: Global gridded soil information based
on machine learning. PLoS One 12, e0169748.
https://doi.org/10.1371/journal.
.
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, M.G., Gonzalez, M.R., 2014.
SoilGrids1km
—
Global Soil Information based on Automated Mapping. PLoS One 9,
e105992.
https://doi.org/10.1371/journal.pone.0105992
Heung, B., Ho, H.C., Zhang, J., Knudby, A., Bulmer, C.E., Schmidt, M.G., 2016. An
overview and comparison of machine-learning techniques for classification purposes
in digital soil mapping. Geoderma 265, 62
–
77.
.
Heuvelink, G.B.M., Kros, J., Reinds, G.J., De Vries, W., 2016. Geostatistical prediction
and simulation of European soil property maps. Geoderma Reg. 7, 201
–
doi.org/10.1016/j.geodrs.2016.04.002
.
Heuvelink, G.B.M., Webster, R., 2001. Modelling soil variation: past, present, and future.
Geoderma, Developments and Trends in Soil Science 100, 269
–
.
Hong, D., Zhang, B., Li, X., Li, Y., Li, C., Yao, J., Yokoya, N., Li, H., Ghamisi, P., Jia, X.,
Plaza, A., Gamba, P., Benediktsson, J.A., Chanussot, J., 2024. SpectralGPT: Spectral
Remote Sensing Foundation Model. IEEE Trans. Pattern Anal. Mach. Intell. 46,
5227
–
5244.
https://doi.org/10.1109/TPAMI.2024.3362475
Huang, H., Yang, L., Zhang, L., Pu, Y., Yang, C., Wu, Q., Cai, Y., Shen, F., Zhou, C., 2022.
A review on digital mapping of soil carbon in cropland: progress, challenge, and
prospect. Environ. Res. Lett. 17, 123004.
https://doi.org/10.1088/1748-9326/
Hudson, B.D., 1992. The Soil Survey as Paradigm-based Science. Soil Sci. Soc. Am. J. 56,
836
–
841.
https://doi.org/10.2136/sssaj1992.03615995005600030027x
Ivushkin, K., Bartholomeus, H., Bregt, A.K., Pulatov, A., Kempen, B., De Sousa, L., 2019.
Global mapping of soil salinity change. Remote Sens. Environ. 231, 111260.
doi.org/10.1016/j.rse.2019.111260
.
Jenny, H., 1941. Factors of Soil Formation: a System of Quantitative Pedology. Agron. J.
33, 857
–
858.
https://doi.org/10.2134/agronj1941.00021962003300090016x
.
Jensen, K.H., Illangasekare, T.H., 2011. HOBE: a Hydrological Observatory. Vadose Zone
J. 10, 1
–
7.
https://doi.org/10.2136/vzj2011.0006
Kingma, D.P., Ba, J., 2017. Adam: A Method for Stochastic Optimization. https://doi.
org/10.48550/arXiv.1412.6980.
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521, 436
–
444.
Lilburne, L., Helfenstein, A., Heuvelink, G.B.M., Eger, A., 2024. Interpreting and
evaluating digital soil mapping prediction uncertainty: a case study using texture
from SoilGrids. Geoderma 450, 117052.
.
Liu, F., Wu, H., Zhao, Y., Li, D., Yang, J.-L., Song, X., Shi, Z., Zhu, A.-X., Zhang, G.-L.,
2022. Mapping high resolution National Soil Information Grids of China. Science
Bulletin 67, 328
–
340.
https://doi.org/10.1016/j.scib.2021.10.013
.
Ma, Y., Minasny, B., Malone, B.P., Mcbratney, A.B., 2019. Pedology and digital soil
mapping (DSM). Eur. J. Soil Sci. 70, 216
–
https://doi.org/10.1111/ejss.12790
.
Ma, Y., Minasny, B., Wu, C., 2017. Mapping key soil properties to support agricultural
production in Eastern China. Geoderma Reg. 10, 144
–
.
Malone, B.P., McBratney, A.B., Minasny, B., Laslett, G.M., 2009. Mapping continuous
depth functions of soil carbon storage and available water capacity. Geoderma 154,
138
–
152.
https://doi.org/10.1016/j.geoderma.2009.10.007
.
McBratney, A.B., Mendonça Santos, M.L., Minasny, B., 2003. On digital soil mapping.
Geoderma 117, 3
–
52.
https://doi.org/10.1016/S0016-7061(03)00223-4
.
Meng, X., Bao, Y., Luo, C., Zhang, X., Liu, H., 2024a. SOC content of global Mollisols at a
30 m spatial resolution from 1984 to 2021 generated by the novel ML-CNN
prediction model. Remote Sens. Environ. 300, 113911.
Meng, X., Bao, Y., Luo, C., Zhang, X., Liu, H., 2024b. A new methodology for establishing
an SOC content prediction model that is spatiotemporally transferable at
multidecadal and intercontinental scales. ISPRS J. Photogramm. Remote Sens. 218,
531
–
550.
https://doi.org/10.1016/j.isprsjprs.2024.09.038
Meng, X., Bao, Y., Wang, Y., Zhang, X., Liu, H., 2022. An advanced soil organic carbon
content prediction model via fused temporal-spatial-spectral (TSS) information
based on machine learning and deep learning algorithms. Remote Sens. Environ.
280, 113166.
https://doi.org/10.1016/j.rse.2022.113166
.
Meyer, H., Pebesma, E., 2021. Predicting into unknown space? estimating the area of
applicability of spatial prediction models. Methods Ecol. Evol. 12, 1620
–
1633.
https://doi.org/10.1111/2041-210X.13650
Minasny, B., McBratney, A.B., Malone, B.P., Wheeler, I., 2013. Digital Mapping of Soil
Carbon, in: Advances in Agronomy. Elsevier, pp. 1
–
47.
.
L. Zhang et al.
Geoderma
461
(2025)
117466
15
Minasny, B., McBratney, Alex.B., 2016. Digital soil mapping: A brief history and some
lessons. Geoderma, Soil mapping, classification, and modelling: history and future
directions 264, 301
–
311.
https://doi.org/10.1016/j.geoderma.2015.07.017
Moor, M., Banerjee, O., Abad, Z.S.H., Krumholz, H.M., Leskovec, J., Topol, E.J.,
Rajpurkar, P., 2023. Foundation models for generalist medical artificial intelligence.
Nature 616, 259
–
265.
https://doi.org/10.1038/s41586-023-05881-4
.
Mulder, V.L., de Bruin, S., Schaepman, M.E., Mayr, T.R., 2011. The use of remote sensing
in soil and terrain mapping
—
a review. Geoderma 162, 1
–
19.
10.1016/j.geoderma.2010.12.018
.
Nussbaum, M., Vogel, S., Oechslin, S., Tanner, S., Burgos, S., 2023. Smoothed predicted
distributions in digital soil mapping
–
a comprehensive comparative study to predict
soil texture for irrigation (No. EGU23-5543). Presented at the EGU23, Copernicus
Meetings. https://doi.org/10.5194/egusphere-egu23-5543.
Orgiazzi, A., Ballabio, C., Panagos, P., Jones, A., Fern
´
andez-Ugalde, O., 2018. LUCAS
Soil, the largest expandable soil dataset for Europe: a review. Eur. J. Soil Sci. 69,
140
–
153.
https://doi.org/10.1111/ejss.12499
Peng, Y., Xiong, X., Adhikari, K., Knadel, M., Grunwald, S., Greve, M.H., 2015. Modeling
Soil Organic Carbon at Regional Scale by Combining Multi-Spectral Images with
Laboratory Spectra. PLoS One 10, e0142295.
https://doi.org/10.1371/journal.
Poggio, L., de Sousa, L.M., Batjes, N.H., Heuvelink, G.B.M., Kempen, B., Ribeiro, E.,
Rossiter, D., 2021. SoilGrids 2.0: producing soil information for the globe with
quantified spatial uncertainty. Soil 7, 217
–
240.
https://doi.org/10.5194/soil-7-217-
Sanchez, P.A., Ahamed, S., Carr
´
e, F., Hartemink, A.E., Hempel, J., Huising, J.,
Lagacherie, P., McBratney, A.B., McKenzie, N.J., Mendonça-Santos, M. de L.,
Minasny, B., Montanarella, L., Okoth, P., Palm, C.A., Sachs, J.D., Shepherd, K.D.,
Vågen, T.-G., Vanlauwe, B., Walsh, M.G., Winowiecki, L.A., Zhang, G.-L., 2009.
Digital Soil Map of the World. Science 325, 680
–
681. https://doi.org/10.1126/
science.1175084.
Sanderman, J., Hengl, T., Fiske, G.J., 2017. Soil carbon debt of 12,000 years of human
land use. Proc. Natl. Acad. Sci. 114, 9575
–
9580.
.
Schaaf, C., Wang, Z., 2021. MODIS/Terra
+
Aqua BRDF/Albedo Nadir BRDF Adjusted Ref
Daily L3 Global - 500m V061. https://doi.org/10.5067/MODIS/MCD43A4.061.
Schmidt, M.W.I., Torn, M.S., Abiven, S., Dittmar, T., Guggenberger, G., Janssens, I.A.,
Kleber, M., K
¨
ogel-Knabner, I., Lehmann, J., Manning, D.A.C., Nannipieri, P.,
Rasse, D.P., Weiner, S., Trumbore, S.E., 2011. Persistence of soil organic matter as an
ecosystem property. Nature 478, 49
–
56.
https://doi.org/10.1038/nature10386
Shen, Z., Ramirez-Lopez, L., Behrens, T., Cui, L., Zhang, M., Walden, L., Wetterlind, J.,
Shi, Z., Sudduth, K.A., Baumann, P., Song, Y., Catambay, K., Viscarra Rossel, R.A.,
2022. Deep transfer learning of global spectra for local soil carbon monitoring. ISPRS
J. Photogramm. Remote Sens. 188, 190
–
200.
.
Tan, T., Genova, G., Heuvelink, G.B.M., Lehmann, J., Poggio, L., Woolf, D., You, F., 2024.
Importance of Terrain and climate for predicting Soil Organic Carbon is Highly
Variable across local to Continental Scales. Environ. Sci. Technol. 58, 11492
–
11503.
https://doi.org/10.1021/acs.est.4c01172
.
Tiessen, H., Cuevas, E., Chacon, P., 1994. The role of soil organic matter in sustaining soil
fertility. Nature 371, 783
–
785.
https://doi.org/10.1038/371783a0
.
Viscarra Rossel, R.A., Minasny, B., Roudier, P., McBratney, A.B., 2006a. Colour space
models for soil science. Geoderma 133, 320
–
Viscarra Rossel, R.A., Shen, Z., Ramirez Lopez, L., Behrens, T., Shi, Z., Wetterlind, J.,
Sudduth, K.A., Stenberg, B., Guerrero, C., Gholizadeh, A., Ben-Dor, E., St Luce, M.,
Orellano, C., 2024. An imperative for soil spectroscopic modelling is to think global
but fit local with transfer learning. Earth Sci. Rev. 254, 104797.
10.1016/j.earscirev.2024.104797
.
Viscarra Rossel, R.A., Walvoort, D.J.J., McBratney, A.B., Janik, L.J., Skjemstad, J.O.,
2006b. Visible, near infrared, mid infrared or combined diffuse reflectance
spectroscopy for simultaneous assessment of various soil properties. Geoderma 131,
59
–
75.
https://doi.org/10.1016/j.geoderma.2005.03.007
Wadoux, A.M.J.C., 2019. Using deep learning for multivariate mapping of soil with
quantified uncertainty. Geoderma 351, 59
–
70.
.
Wadoux, A.-M.-J.-C., Heuvelink, G.B.M., Lark, R.M., Lagacherie, P., Bouma, J.,
Mulder, V.L., Libohova, Z., Yang, L., McBratney, A.B., 2021. Ten challenges for the
future of pedometrics. Geoderma 401, 115155.
.
Wadoux, A.-M.-J.-C., Minasny, B., Mcbratney, A.B., 2020. Machine learning for digital
soil mapping: applications, challenges and suggested solutions. Earth Sci. Rev. 210,
103359.
https://doi.org/10.1016/j.earscirev.2020.103359
.
Wang, B., Gray, J.M., Waters, C.M., Rajin Anwar, M., Orgill, S.E., Cowie, A.L., Feng, P., Li
Liu, D., 2022. Modelling and mapping soil organic carbon stocks under future
climate change in south-eastern Australia. Geoderma 405, 115442.
10.1016/j.geoderma.2021.115442
Wang, X., Li, S., Wang, L., Zheng, M., Wang, Z., Song, K., 2023. Effects of cropland
reclamation on soil organic carbon in China
’
s black soil region over the past 35
years. Glob. Chang. Biol. 29, 5460
–
5477.
https://doi.org/10.1111/gcb.16833
Wu, T., He, S., Liu, J., Sun, S., Liu, K., Han, Q.-L., Tang, Y., 2023. A Brief Overview of
ChatGPT: the history, Status Quo and potential Future Development. IEEE/CAA J.
Autom. Sin. 10, 1122
–
https://doi.org/10.1109/JAS.2023.123618
Yamazaki, D., Ikeshima, D., Tawatari, R., Yamaguchi, T., O
’
Loughlin, F., Neal, J.C.,
Sampson, C.C., Kanae, S., Bates, P.D., 2017. A high-accuracy map of global terrain
elevations. Geophys. Res. Lett. 44, 5844
–
5853.
Yang, L., Cai, Y., Zhang, L., Guo, M., Li, A., Zhou, C., 2021a. A deep learning method to
predict soil organic carbon content at a regional scale using satellite-based
phenology variables. Int. J. Appl. Earth Obs. Geoinf. 102, 102428.
Yang, L., Li, X., Yang, Q., Zhang, L., Zhang, S., Wu, S., Zhou, C., 2021b. Extracting
knowledge from legacy maps to delineate eco-geographical regions. Int. J. Geogr.
Inf. Sci. 35, 250
–
272.
https://doi.org/10.1080/13658816.2020.1806284
.
Zhang, L., Cai, Y., Huang, H., Li, A., Yang, L., Zhou, C., 2022a. A CNN-LSTM model for
soil organic carbon content prediction with long time series of MODIS-based
phenological variables. Remote Sens. (Basel) 14, 4441.
.
Zhang, L., Heuvelink, G.B.M., Mulder, V.L., Chen, S., Deng, X., Yang, L., 2024. Using
process-oriented model output to enhance machine learning-based soil organic
carbon prediction in space and time. Sci. Total Environ. 922, 170778.
org/10.1016/j.scitotenv.2024.170778
.
Zhang, L., Yang, L., Cai, Y., Huang, H., Shi, J., Zhou, C., 2022b. A multiple soil properties
oriented representative sampling strategy for digital soil mapping. Geoderma 406,
115531.
https://doi.org/10.1016/j.geoderma.2021.115531
.
Zhang, L., Yang, L., Crowther, T.W., Zohner, C.M., Doetterl, S., Heuvelink, G.B.M.,
Wadoux, A.-M.-J.-C., Zhu, A.-X., Pu, Y., Shen, F., Ma, H., Zou, Y., Zhou, C., 2025.
Mapping global distributions, environmental controls, and uncertainties of apparent
topsoil and subsoil organic carbon turnover times. Earth Syst. Sci. Data 17,
2605
–
2623.
https://doi.org/10.5194/essd-17-2605-2025
Zhang, L., Yang, L., Ma, T., Shen, F., Cai, Y., Zhou, C., 2021. A self-training semi-
supervised machine learning method for predictive mapping of soil classes with
limited sample data. Geoderma 384, 114809.
.
Zhu, A., Lu, G., Liu, J., Qin, C., Zhou, C., 2018. Spatial prediction based on Third Law of
Geography. Ann. GIS 24, 225
–
Zhu, A.-X., Hudson, B., Burt, J., Lubich, K., Simonson, D., 2001. Soil Mapping using GIS,
Expert Knowledge, and Fuzzy Logic. Soil Sci. Soc. Am. J. 65, 1463
–
org/10.2136/sssaj2001.6551463x
.
Zhu, A.X., Liu, J., Du, F., Zhang, S.J., Qin, C.Z., Burt, J., Behrens, T., Scholten, T., 2015.
Predictive soil mapping with limited sample data. Eur. J. Soil Sci. 66, 535
–
547.
https://doi.org/10.1111/ejss.12244
.
L. Zhang et al.
Geoderma
461
(2025)
117466
16