Zhao et al. (2026). The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world farms in Northeast China

Rapid #: -27733242

CROSS REF ID:

36821144720001865

LENDER:

KWC (University of the Cumberlands) :: Main Library

BORROWER:

ORE (Oregon State University) :: Main Library

TYPE:

Article CC:CCG

JOURNAL TITLE:

Nature food

USER JOURNAL TITLE:

Nature Food

ARTICLE TITLE:

The beneficial relation between soil organic carbon and maize yield in field trials does not
translate to real-world farms in Northeast China

ARTICLE AUTHOR:

Zhao, Shuaixiang

VOLUME:

7

ISSUE:

8

MONTH:

YEAR:

2026-08-01

PAGES:

782 - 791

ISSN:

2662-1355

OCLC #:

Processed by RapidX:

10/20/2026 7:54:40 AM

This material may be protected by copyright law (Title 17 U.S. Code)

|

|

|

User Name: =

Date and Time: = 2026-10-20

Job Number: = 299360716

Documents (1)

Client/Matter:

-None-

Search Terms:

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to

real-world farms in Northeast China

Search Type:

NaturalAnd

Content Type

Narrowed by

news

-None-

1. The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world
farms in Northeast China

Page 1 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

The beneficial relation between soil organic carbon and maize yield in field

trials does not translate to real-world farms in Northeast China

Nature Food

August 2026

Section: Pg. 782-791; Vol. 7; No. 8; ISSN: 2662-1355

Length: 6767 words

Byline: Shuaixiang Zhao;

1

Susanne Schmidt;

2

Zhi Quan;

3

Fenglan Li;

4

Shuqi Qin;

1

Yunting Fang;

31

Feng Zhou;

1552

Body

Abstract

1

Institute of Carbon Neutrality, Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Sino-

France Institute of Earth System Science, Peking University, Beijing, , China, 0000 0001 2256 9319,
grid.11135.37,

https://ror.org/02v51f717

2

School of Agriculture and Food Sustainability, The University of Queensland, Brisbane, Queensland, Australia, 0000 0000 9320

7537, grid.1003.2,

https://ror.org/00rqy9422

3

Institute of Applied Ecology, Chinese Academy of Sciences, Shenyang, , China, 0000 0001 1957 3309,

grid.9227.e,

https://ror.org/034t30j35

4

College

of

Life

Sciences,

Northeast

Agricultural

University,

Harbin,

,

China,

0000

0004

1760

1136,

grid.412243.2,

https://ror.org/0515nd386

1

Institute of Carbon Neutrality, Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Sino-

France Institute of Earth System Science, Peking University, Beijing, , China, 0000 0001 2256 9319,
grid.11135.37,

https://ror.org/02v51f717

3

Institute of Applied Ecology, Chinese Academy of Sciences, Shenyang, , China, 0000 0001 1957 3309,

grid.9227.e,

https://ror.org/034t30j35

fangyt@iae.ac.cn

1

Institute of Carbon Neutrality, Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Sino-

France Institute of Earth System Science, Peking University, Beijing, , China, 0000 0001 2256 9319,
grid.11135.37,

https://ror.org/02v51f717

5

, Southwest United Graduate School, Kunming, , China, , ,

zhouf@pku.edu.cn

Page 2 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

Carbon  sequestration  in  agricultural  soils  is  an  economically  viable  pathway  to  achieve  climate  and  food  co-
benefits,  but  it  remains  unclear  whether  the  positive  soil  organic  carbon  (SOC)–yield  relationship  observed  in
field trials applies to real-world farms. Here we analysed maize yield responses to SOC on 1,246 field trials and
192  farms  in  Northeast  China  using  multiple  causal  inference  approaches.  We  found  that  an  increase  in  SOC
consistently  enhanced  maize  yield  in  field  trials  but  had  no  detectable  effect  in  farms.  Field  trials  showed
significant yield increases of 8–13% up to a threshold of 30–40 g SOC

kg−1

soil. On farms, however, yields rose

by only up to 5% with increasing SOC to a threshold of 10.0 g SOC

kg−1

soil and then decreased to around the

initial  yield.  The  absence  of  a  positive  yield–SOC  response  in  real-world  farms  was  mainly  attributable  to
excessive  nitrogen  fertilizer  applications,  along  with  limited  farmer  knowledge  and  inadequate  infrastructure
hindering  effective  adoption  of  soil  conservation  practices.  Our  findings  underscore  the  need  to  integrate
fertilization, soil conservation strategies and farmer support to fully realize the potential of SOC on yield.

Soil organic carbon (SOC) is widely recognized for its potential to enhance agricultural productivity and contribute
to climate mitigation, yet its benefits in real-world farming contexts remain uncertain. This study reveals that while
increased SOC significantly boosts maize yields in controlled field trials, it has negligible effects on actual farm
yields in Northeast China, primarily due to excessive nitrogen fertilization and inadequate adoption of soil
conservation practices.

Main

Achieving  net-zero-emissions  food  systems  is  central  to  the  Sustainable  Development  Goals1.  Nearly  half  of  all
countries  have  adopted  carbon  sequestration  in  agricultural  soils  as  an  economically  viable  pathway  for  climate
mitigation with the co-benefit of increasing soil organic carbon (SOC)2,3. Increasing SOC improves soil structure,
water retention, nutrient supply and microbial function4, all of which, indicated by multiple pieces of evidence from
on-station  or  on-farm  field  trials,  can  enhance  crop  production5,  6–7.  However,  much  less  is  known  about  real-
world  farms  managed  by  farmers  using  local  practice  with  limited  scientific  guidance8.  This  knowledge  gap
motivated  us  to  examine  whether  the  beneficial  relationship  between  SOC  and  yield  observed  in  field  trials
translates to real-world farms and, if not, to identify the limiting factors.

Unlike  the  field  trials,  which  control  many  variables  to  isolate  the  SOC–yield  relationship,  farms  represent  real-
world  conditions  that  may  obscure  the  beneficial  effects  of  SOC.  Suboptimal  fertilization,  heterogeneous  field
conditions, and limitations in farmers’ knowledge and access to resources may further hinder the effective adoption
of  soil  conservation  practices,  such  as  manure  use,  straw  return  and  no-till,  for  enhancing  SOC  or  crop  yield9.
These  management  practices,  designed  to  enhance  SOC,  can  also  affect  crop  yields,  potentially  disordering  the
SOC–yield relationship10. This is more critical in regions where farmers are encouraged to enhance SOC because
its  depletion  has  caused  severe  soil  degradation,  threatening  crop  productivity  and  long-term  agricultural
sustainability.

Here  we  focus  on  the  maize  crop  in  Northeast  China  because  this  region  accounts  for  a  quarter  of  China’s
cropland  and  40%  of  national  maize  production11,  yet  it  is  the  only  region  in  China  experiencing  a  continuous
decline  in  SOC  over  the  past  6  decades12.  The  Chinese  government  has  recently  launched  a  series  of  policy
interventions to enhance SOC for improving crop productivity11. Clarifying the SOC–yield relationship on farms in
Northeast  China  is  crucial  for  optimizing  such  interventions,  especially  considering  that  maize  yields  have
stagnated  in  recent  years13.  However,  this  is  particularly  challenging  due  not  only  to  the  lack  of  paired
measurements of SOC and yield from large-scale surveys of real-world farms8 but also to the difficulty of isolating
the  causal  effect  of  SOC  on  maize  yield,  which  is  often  confounded  by  soil,  climate  and  management
practices5,10.

To  address  these  challenges,  we  conducted  a  large-scale  survey  of  192  real-world  farms  using  multiple  causal
inference  approaches  to  quantify  the  SOC–yield  relationship  in  Northeast  China,  which  spans  a  wide  range  of
soils,  climates  and  management  practices  (Supplementary  Fig.  1  and  Supplementary  Table  1).  To  examine
differences  in  the  SOC–yield  relationship,  the  same  analyses  were  further  performed  for  field  trials,  including
1,030 on-farm and 216 on-station trials in this region. These analyses allowed us to address three key questions.

Page 3 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

What is the magnitude of the effect of SOC on maize yield in Northeast China? Does the effect of SOC on maize
yield differ between farmer-managed fields and controlled field trials? What are the key mechanisms driving
maize yield responses to SOC?

Results and discussion

Divergent SOC–yield relationships

We used causal forest (CF) and double machine learning (DML), which rely on distinct estimation principles, to
quantify the effects of SOC on maize yields in both farms and field trials. These approaches yielded consistent
causal effect estimates, revealing that increasing SOC consistently enhanced maize yield in field trials but had no
detectable effect on farms across Northeast China (Fig. 1b). For farms, the mean causal effects were

−62.3

kg

ha−1

(CF) and

−56.0

kg

ha−1

(DML) per unit increase in SOC (hereafter referring to 1 g

kg−1),

with 95%

confidence intervals (CIs) of

−183.2

to 48.7 (CF) and

−137.7

to 25.8 (DML). The mean effects were 40.7 (21.1–

60.3, CF) and 21.4

(−1.4

to 44.1, DML) for on-farm trials and 118.2 (13.0–223.2, CF) and 204.3 (52.6–356.0, DML)

for on-station trials. These causal results were confirmed by farm and field trial data from regions with comparable
soil and climate conditions (Supplementary Fig. 2). When further accounting for potential biases arising from
discrepancies between farms and field trial data, as well as modelling uncertainties, these causal conclusions
remained robust (Supplementary Fig. 3).

Fig. 1

Geographic distribution of the studied data and the relationship between SOC and maize yield across
Northeast China.

a,  Map  of  China  with  Northeast  China  in  grey  and  the  distribution  of  sites.  Coloured  points  represent  data  from
farms  (blue,  n  =  192),  on-farm  trials  (yellow,  n  =  1,030)  and  on-station  trials  (brown,  n  =  216).  Farms  were
surveyed  from  three  counties:  Changtu  (CT),  Lishu  (LS)  and  Gannan  (GN).  b,  Causal  effects  of  SOC  on  maize
yields. Two methods, CF and DML, were applied. Data are represented as mean ± 95% confidence interval of the
causal effects across farms, on-farm trials and on-station trials. Error bars not overlapping 0 indicate P < 0.05. c,
Relationships  between  SOC  and  maize  yield.  Maize  yield  was  normalized  to  the  yield  with  the  lowest  SOC
content,  which  was  9.2  Mg

ha−1

(farms), 9.3 Mg

ha−1

(on-farm trials) and 10.3 Mg

ha−1

(on-station trials). The

fitted regression lines with shaded areas represent the partial dependence of the RF model for SOC–yield
relationships, and the shaded regions correspond to the 95% CIs. The bars at the bottom of each panel represent
the distribution of SOC. Base map of the China and provincial administrative boundary was obtained from the
Resource and Environmental Science Data Platform of China (

https://www.resdc.cn/DOI/DOI.aspx?DOIID=122

).

Source data

Partial dependence analysis of random forest (RF) models revealed differences in the SOC–yield relationship
between field trials and farms. Field trials showed 8% and 13% higher maize yield with increasing SOC up to
thresholds of 30.1 g

kg−1

for on-station field trials and 39.5 g

kg−1

for on-farm field trials (Fig. 1c). In contrast,

maize yield on farms increased by only up to 5% as SOC reached 10.0 g

kg−1

before decreasing to around the

initial  yield.  The  divergent  SOC–yield  relationships  were  supported  by  Shapley  additive  explanations  (SHAP)
analysis.  A  general  increasing  trend  between  SOC  and  its  SHAP  values  was  observed  in  both  on-farm  and  on-
station  trials  (Fig.  2e,f,h,i),  indicating  an  overall  positive  effect  of  SOC  on  maize  yield  across  the  observed  SOC
range. However, SHAP values for farms showed that the beneficial effect of SOC on maize yield first increased
and then decreased with increasing SOC content (Fig. 2b,c).

Fig. 2

Factors influencing maize yield.

Page 4 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

a,d,g, Mean absolute SHAP values showing the relative importance of variables for maize yield on farms (a), in
on-farm trials (d) and in on-station trials (g). b,e,h, Beeswarm plots of SHAP values for farms (b), on-farm trials (e)
and  on-station  trials  (h).  Dots  represent  the  SHAP  value  of  each  factor  across  all  individual  observations.  Point
colour  represents  the  normalized  feature  value,  with  red  indicating  high  values  and  blue  indicating  low  values.
Higher SHAP values indicate higher predicted yields, and lower SHAP values indicate lower predicted yields. The
labels for the axes in b, e and h are the same as those in a, d and g, respectively. c,f,i, SHAP dependence plot
showing the relationship between SOC and its SHAP values for farms (c), on-farm trials (f) and on-station trials (i).
SHAP values were calculated using RF models. SHAP values (y axis) indicate the marginal contribution of SOC to
maize  yield  (Mg

ha−1).

Asterisks indicate statistically significant effects from two-sided permutation tests (**P <

0.01 and *P < 0.05). NS, nonsignificant.

Source data

RF–SHAP analyses also identified divergent determinants of maize yield variability between farms and field
trials. Nitrogen (N) fertilizer input was the top driver of yield variation in field trials (Fig. 2d,g). Maize yield
increased with N input and plateaued at ~150 kg N

ha−1

for both on-farm and on-station trials, but yield on farms

did  not  change  with  N  input  (Supplementary  Fig.  4).  In  contrast,  growing  period  rainfall  (GPR)  dominated  yield
variations  on  farms  (Fig.  2a),  with  yield  being  positively  correlated  with  GPR  (Supplementary  Fig.  4).  High-SOC
soils  in  farms  were  more  prominent  in  regions  with  lower  GPR  (Supplementary  Fig.  5),  which  may  explain  the
lower yields in these regions.

Drivers of SOC–yield relationships

To  further  explore  the  drivers  of  SOC–yield  response,  we  considered  ten  potential  factors  influencing  this
relationship,  including  three  fertilizer  input  variables  (N,  phosphorus  (P2O5)  and  potassium  (K2O)),  two  soil
variables  (clay  content  and  pH  level),  two  climatic  variables  (GPR,  growing  period  temperature  (GTE))  and  three
soil management variables (manure use, straw return and no-till). A combined approach integrating sliding-window
analysis with partial correlation enabled us to evaluate how each factor regulates the impact of SOC on maize yield
(Methods).

The results showed that higher fertilizer rates reduced the partial correlation between SOC and yield (Fig. 3a–c).
This  pattern  was  confirmed  in  both  farms  and  field  trials,  as  well  as  for  almost  all  N,  P2O5  (except  for  on-farm
trials)  and  K2O  fertilizers.  Data  from  three  long-term  (16–38  years)  field  trial  sites  also  showed  a  weaker  SOC–
yield correlation under a higher N rate than under a lower N rate (Supplementary Fig. 6). In line with this result, a
controlled greenhouse study of spring wheat also showed that N input generated greater yield benefits in low-SOC
soils,  owing  to  higher  net  N  mineralization  in  high-SOC  soils14.  Excessive  fertilization  may  further  suppress  root
growth, thereby reducing the uptake of water and nutrients15,16. This may help explain why SOC enhanced yield
in  field  trials  but  not  on  farms,  which  received  50%  more  fertilizer  (225  kg  N

ha−1

versus 150 kg N

ha−1

on

average) and had approximately twice the available soil nutrient content of field trials (Supplementary Table 1).

Fig. 3

The SOC–yield partial correlation response to environmental and management variables.

a–h, Effects of nitrogen (a), phosphorus (b) and potassium (c) fertilization, soil clay content (d), soil pH (e), GPR
(f), GTE (g) and soil conservation practices (h), including manure use, straw return and no-till. Coloured lines
represent values for farms (blue), on-farm trials (yellow) and on-station trials (brown). Regression lines show linear
(a,b,f,g) and quadratic (d,e) models, with error bands indicating 95% CIs. Model performance is represented by the
regression coefficient (R2) and the P value, with significant regressions (P < 0.05, two-sided t-test). Model
equations are provided in Supplementary Table 5.

Source data

Page 5 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

The  effects  of  soil  clay  content  and  soil  pH  value  on  the  SOC–yield  relationship  were  nonlinear  (Fig.  3d,e).
Positive SOC–yield partial correlations occurred often in sandy soils (clay content <15%) and clay-rich soils (clay
content >35%)17, mainly because of improved soil structure and nutrient buffering capacity in these soils6. Positive
SOC–yield partial correlations also occurred often in soil pH >5.5, while a negative correlation emerged at pH <5.5,
suggesting  that  SOC  provides  fewer  yield  benefits  in  acidic  soils.  This  contrasts  with  previous  research  showing
strong  yield  benefits  of  biochar  amendment  in  acidic  soils,  primarily  through  soil  liming18,19.  This  further
underscores the detrimental impact of soil acidity on crop production by inhibiting root function involved in nutrient
cycling  and  reducing  the  bioavailability  of  essential  nutrients  (for  example,  phosphorus  and  micronutrients)  in  the
rhizosphere20,21.  High  N  inputs,  which  often  co-occurred  with  low  soil  pH  (Supplementary  Fig.  7),  may  further
weaken  the  SOC-derived  yield  benefits.  These  results  highlight  the  need  to  improve  soil  pH  on  farms  to
strengthen the SOC–yield relationship, as farms showed lower pH than field trials (average 5.6 versus 6.3–6.6;
Supplementary Table 1), probably owing to prolonged high N fertilization20.

The  SOC–yield  partial  correlation  declined  with  increasing  GPR  and  GTE  on  farms  (Fig.  3f,g),  indicating  greater
SOC-derived yield benefits under dry and/or cold climates. This is primarily attributed to the positive effects of SOC
on water-holding capacity in low-rainfall regions22 and on soil nutrient availability in cold regions6,23. In contrast,
the  SOC–yield  partial  correlation  in  on-farm  trials  showed  no  responses  to  GPR  or  GTE.  While  the  SOC–yield
partial  correlation  response  to  GTE  in  on-station  trials  mirrored  that  in  on-farm  trials,  the  response  to  GPR  was
reversed. These results suggest that the climate sensitivity of the SOC–yield relationships differs between farms
and field trials. One possible explanation is contrasting soil fertility. On-farm trials had higher initial soil SOC (17.8
g C

kg−1)

than farms and on-station trials (14.5 g C

kg−1

and 15.1 g C

kg−1,

respectively; Supplementary Table

1). This enhances soil water and nutrient regulation and reduces the climate resilience of on-farm trials relative to
the other two. Another explanation is that crop yields in many field trials are often estimated from small-area
sampling, which can underestimate climate impacts, such as yield losses from waterlogging in low-lying areas24. In
contrast, farm yields reflect the overall harvest of fields and therefore better capture the impacts of climate.

Soil  conservation  practices  (manure  use,  no-till  and  straw  return)  also  affect  the  SOC–yield  relationship,  with
marked  differences  between  farms  and  field  trials  (Fig.  3h).  Compared  with  traditional  practices,  manure
application  and  no-till  increased  the  SOC–yield  partial  correlation  in  on-station  trials  but  had  minimal  effects  on
farms.  This  contrast  probably  stems  from  the  quality  of  implementation  of  these  practices.  Manure  application  in
on-station  trials  typically  follows  an  optimized  method  with  a  knowledge-based  rate,  thereby  resulting  in  greater
SOC-derived  yield  benefits.  In  contrast,  farms  often  lack  the  machinery  and  knowledge,  leading  to  improper
manure  application,  with  43%  of  manure-using  farmers  reporting  no  yield  gains  (Supplementary  Fig.  8).  Straw
return  is  an  essential  part  of  no-till  because  it  improves  soil  structure  and  increases  water-use  efficiency25.
However, among farmers adopting no-till, only 33% practiced both straw return and no-till, compared with 83% in
on-station  trials  (Supplementary  Fig.  9).  This  discrepancy  helps  explain  why  63%  of  no-till  farmers  reported  no
obvious yield benefits (Supplementary Fig. 10).

Implications and perspectives

Our survey demonstrates that the beneficial relationship between SOC and maize yield observed in field trials
does not translate to real-world farms in Northeast China. Our findings align with those of a regional farm
survey in Europe, which also showed a weak SOC–yield association but relied on farmers’ self-reported SOC26.
Our results contrast with a farm survey in the USA, which identified a positive SOC–yield relationship among
farmers practicing tight nutrient management8. This finding reinforces our interpretation: when fertilizer is applied
judiciously, farms can capitalize on the yield-promoting benefits of SOC.

The absence of yield benefits from SOC on farms in Northeast China can be attributed to high fertilizer inputs
(Fig. 4). This suggests that the real-world yield benefits of SOC are negligible in highly fertilized regions such as
China. This is probably different in areas where fertilizer underuse is common, such as sub-Saharan Africa6. In
such systems, SOC accrual via fertilizer input combined with organic amendments (for example, crop residues and
manure) remains highly beneficial for crop productivity. Irrespective of the region, achieving sustainable SOC–crop

Page 6 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

yield interplay requires context-specific strategies that balance trade-offs between SOC sequestration, nutrient
availability and broader ecosystem functions10.

Fig. 4

Conceptual summary of the divergent SOC–yield relationship between farms and field trials.

The positive SOC–yield link characterized field trials but not farms. Soil conservation practices have synergistic
effects on SOC and crop yield in field trials, in contrast to farms. Arrow thickness indicates the strength of the
effects shown in the explanatory panel. Icons were adapted from the iSlide plugin icon library
(

https://www.islide.cc/

), which is available for free use under the CC0 license.

Source data

Our analysis also suggests that government-recommended soil practices may fail to deliver the expected SOC-
derived gains on farms because of management deficiencies in real-world practices across Northeast China (Fig.
4). For example, although combining no-till with straw return is recommended to enhance SOC-derived yield25, low
temperatures in Northeast China may hinder straw decomposition and seedling emergence, discouraging farmers
from adopting both practices together. In our survey, more than 70% of farmers either burned or removed straw
rather than returning it to the soil, as suggested by the government (Supplementary Fig. 10).

If  SOC-derived  yield  benefits  are  also  absent  in  most  other  real-world  crops  and  regions,  our  results  have
important scientific and policy implications. First, financial and technical support for farmers should shift from front-
end promotion of SOC accrual to back-end assessment of in-field benefits to identify context-dependent measures
that  jointly  enhance  SOC  and  yield  under  real-world  farming  conditions  while  also  identifying  ineffective  or
adverse ones. Second, our findings raise a critical cross-scale question: how generalizable are the results of small-
plot  controlled  experiments  to  real-world  agriculture?  More  empirical  evidence  is  needed  to  test  the  SOC-
enhancing  measures  in  diverse  agroecosystems.  Third,  SOC  remains  relevant  to  yields,  but  farm  and  research
condition  differences  (fertilization,  management  practices  and  so  on)  mediate  its  benefits.  Achieving  meaningful
yield improvements thus requires integrated measures with coordinated actions across sectors within and beyond
agriculture.

In summary, our findings challenge the prevailing belief of a positive SOC–yield relationship and provide empirical
evidence  from  real-world  farms  where  such  optimization  is  not  achieved.  To  advance,  we  call  for  further
experiments  to  reveal  potential  underlying  mechanisms  that  may  have  been  overlooked.  In  particular,  more
mechanistic studies are needed to elucidate how the quantity and quality of SOC regulate crop yield formation and
stability  through  their  integrated  effects  on  soil  physical  properties  (for  example,  water  retention  and  structure),
chemical  nutrient  dynamics  and  microbial  activity  under  the  complex,  socioeconomically  constrained  settings  of
real-world farming. Such insights are vital for scaling up carbon projects that boost yields while advancing soil
health, climate mitigation and food security.

Methods

Dataset compilation

Farm dataset

The farm dataset was compiled by pairing a farmer survey with soil sampling. The farmer survey was conducted
for  the  major  spring  maize  crop  across  Northeast  China  from  July  to  August  2024.  We  selected  three  major
maize-producing counties (Changtu, Lishu and Gannan) in Northeast China. In each county, 4–6 townships were
randomly  selected,  followed  by  1–5  villages  per  township  and  1–12  farmers  per  village.  Finally,  a  total  of  192
maize farmers from 34 villages across 14 towns, 3 counties and 3 provinces were surveyed (Supplementary Fig. 1
and  Supplementary  Table  2).  Face-to-face  interviews  were  used  to  collect  farm  information.  As  each  farmer
usually managed multiple maize plots with different management practices, our survey focused on the largest plot

Page 7 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

among them. We collected management information on plot area, total fertilizer use, maize yield and soil
management practices (that is, use or no use of manure, tillage and straw management). All information was
recorded for 2023, 1 year before soil sampling. All maize yields were standardized to 14% moisture content. Total
application rates of N, P2O5 and K2O were calculated based on the total quantity of fertilizer applied and its nutrient
content.

Soil samples were collected from the 0–20 cm topsoil of each farmer’s largest maize plot, and locations were
recorded using GPS. Five to six points (5 m apart) were sampled per field, composited into ~1 kg fresh samples
and stored at

−20

°C before further processing. We used a standard method27 to measure soil data of SOC, soil

total  nitrogen  (TN),  available  phosphorus  (AP),  available  potassium  (AK)  concentration,  soil  pH  and  soil  clay
content.  Specifically,  SOC  was  measured  using  the  potassium  dichromate  volumetric  method.  Soil  TN  was
measured  by  the  Kjeldahl  nitrogen  analyser  method.  Soil  AP  was  analysed  by  the  Olsen  method.  Soil  AK  was
measured  with  a  flame  photometer  after  extraction  with  1  M  ammonium  acetate.  Soil  pH  was  measured  using  a
glass electrode in a 1:2.5 soil/water suspension. Soil clay content was measured by the hydrometer method. GPR
and GTE were extracted from the WorldClim database (

www.worldclim.org

) based on sampling coordinates.

Field trial dataset

The field trial dataset consisted of two data sources. One was from on-farm trials conducted from 2005 to 2013
within the National Soil Test and Fertilizer Recommendation Project, which was launched to assess nutrient use
efficiency and optimize fertilizer recommendations across China28. The other was from on-station trials reported in
published studies.

As previous research28 has already provided a complete on-farm trial dataset covering diverse soils and climates
in China, we selected from it according to our specific requirements. In this study, a total of 1,030 site-years of on-
farm  trials  for  spring  maize  across  Northeast  China  were  chosen.  These  trials  were  designed  and  managed  by
local  agricultural  experts  and/or  trained  extension  officers  and  were  implemented  in  on-farm  fields.  For  each  on-
farm  trial,  ten  indicators  were  selected:  soil  organic  matter  (SOM);  AP,  AK  and  pH;  GPR  and  GTE;  total
application rates of N, P2O5 and K2O; and maize yield. Notably, the total application rates of N, P2O5 and K2O
followed optimized fertilizer recommendations developed by local agricultural experts. Soil texture was not included
in the on-farm trial dataset because it was derived from the online Gridded dataset of the Harmonized World Soil
Database  rather  than  laboratory  measurements.  All  these  trials  used  conventional  tillage  without  straw  return  or
manure application. Topsoil (0–20 cm) sampling was consistent with that of our farm dataset. SOM was converted
to  SOC  using  the  simple  equation  SOM/1.724  (ref.  29).  All  yield  data  were  standardized  to  14%  moisture  for
uniform comparison.

We obtained on-station trial data from field experimental studies by searching the ISI Web of Science
(

https://www.webofscience.com/

) and China National Knowledge Infrastructure (

https://www.cnki.net/

) databases.

The search terms were ‘maize’ OR ‘corn’ OR ‘spring maize’ AND ‘soil organic carbon’ OR ‘soil organic matter’
OR ‘SOC’ OR ‘SOM’ AND ‘crop yield’ AND ‘Northeast China’ OR ‘black soil region’. Studies were included if they
met the following criteria: (1) field-based conditions; (2) conducted in Northeast China; and (3) concurrent reporting
of topsoil SOC and SOM content and maize yield. We extracted site coordinates, initial soil properties (SOC/SOM,
TN, AP, AK, pH), fertilizer inputs (N, P2O5  and K2O), soil management practices (manure use, tillage and straw
return), maize yield and soil sampling depth from each paper. To obtain complete data on the above information,
we  excluded  studies  that  reported  only  one-sided  data  (that  is,  articles  reporting  only  SOC  without  other  soil
nutrient  properties).  GPR  and  GTE  of  each  small-plot  trial  were  also  obtained  from  the  WorldClim  database
(

www.worldclim.org

) using the trial site coordinates. Topsoil depth (20 cm) and yield moisture (14%) were

standardized for consistency. As a result, a total of 216 paired SOC–yield records from 35 studies were compiled
for analysis following the Preferred Reporting Items for Systematic reviews guidelines (Supplementary Fig. 11).

Description and comparison of datasets

Page 8 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

We  compared  climate,  soil  and  management  differences  across  the  three  datasets  to  assess  their  comparability
(Supplementary  Table  1).  On-station  trials  had  the  highest  GPR  (520  mm),  followed  by  farms  (487  mm)  and  on-
farm  trials  (433  mm).  On-farm  trials  had  the  highest  GTE  (19.5  °C),  followed  by  on-station  trials  (16.8  °C)  and
farms  (17.0  °C).  Farms  had  relatively  lower  SOC  (14.5  g  C

kg−1)

and soil pH (5.7), but notably higher soil AP

(55.3 mg

kg−1)

and AK (219.2 mg

kg−1)

than on-farm trials (SOC, 17.8 g C

kg−1;

AP, 26.4 mg

kg−1;

AK, 145.3 mg

kg−1)

and on-station trials (SOC, 15.1 g C

kg−1;

AP, 27.9 mg

kg−1;

AK, 156.4 mg

kg−1).

On-farm and on-station

trials had similar total application rates of N (~150 kg

ha−1),

P2O5 (70–80 kg

ha−1)

and K2O (~70 kg

ha−1),

all

significantly lower than those of farms (N, 225 kg

ha−1;

P2O5, 95 kg

ha−1;

K2O, 103 kg

ha−1).

Most farms (>70%)

implemented traditional soil management practices, with no adoption of no-tillage, straw return or manure
application. Maize cultivation in both farms and field trials was entirely rain-fed, with no irrigation applied.

We further quantified the comparability of key climatic and soil variables across the datasets by calculating overlap
coefficients (OVLs), defined as the common area under two probability densities30 (Supplementary Fig. 12) using
the R (version 4.3.2) Package ‘overlapping’31. OVL ranges from 0 (no overlap) to 1 (complete overlap). Specifically,
the  OVL  for  SOC  across  all  datasets  was  0.8–0.9,  indicating  a  relatively  consistent  distribution  of  SOC  data
between  farms  and  field  trials.  For  GTE  and  GPR,  the  similarity  between  farm  and  field  trials  was  moderate,
with  OVLs  of  0.51  and  0.52  (farms  versus  on-farm  trials)  and  0.53  and  0.59  (farms  versus  on-station  trials),
respectively. Similarity in fertilization application rates was lower for N (OVL range, 0.46–0.63) than for P2O5 (0.58–
0.76) and K2O (0.47–0.66). These analyses indicate that the soil and climatic conditions of farms are comparable
to those in on-farm and on-station trials.

Quantifying causal effects of SOC on maize yield

The aim of this study is to determine whether increased SOC leads to improved maize yield. The use of
observational data in farms posed a pervasive challenge for identifying this causal effect, as treatment status may
be correlated with confounding factors. To overcome this inferential hurdle, we used causal machine learning
approaches to estimate treatment effects from observational data32.

We used two approaches to estimate the causal effects of SOC on maize yield in both farms and field trials. The
first  was  CF,  an  adaptation  of  the  classic  RF  algorithm  designed  to  estimate  and  infer  causal  effects  in
observational  studies33.  The  CF,  widely  used  in  the  social  sciences,  is  now  gaining  traction  in  agronomic
research34,35.  Like  RFs,  CF  generates  a  large  number  of  decision  trees,  called  causal  trees.  Each  tree  splits
samples into small subgroups by randomly selecting features and subsets of data. Unlike traditional decision trees,
which  split  to  minimize  prediction  error  in  building  trees,  causal  trees  split  to  maximize  the  variance  in  treatment
effects between the two resulting leaves and make samples more uniform within each subgroup36. For our study,
this means ensuring that samples within a subgroup are as similar as possible in all conditions except SOC, thereby
enabling  us  to  estimate  the  causal  effect  of  SOC  on  yield  within  that  subgroup.  In  this  way,  CF  addresses  the
fundamental problem of causal inference posed by the unobservability of true treatment effects.

The  second  approach  was  DML,  designed  to  overcome  the  challenge  of  evaluating  causal  effects  in  complex
nonlinear  datasets,  especially  when  confounding  variables  are  present37.  DML  estimates  the  causal  effect  of  a
treatment  variable  on  an  outcome  variable  in  two  stages.  In  the  first  stage,  flexible  machine  learning  algorithms
predict  the  treatment  and  outcome  variables  separately,  controlling  for  confounders  and  then  calculate  their
residuals (differences between actual and predicted values). In the second stage, these residuals are incorporated
into a regression model to estimate the causal effect. By accounting for confounders, DML enables more accurate
estimation of the ‘relatively pure’ effect of the treatment on the outcome, supporting causal inference. DML balances
interpretability  and  modelling  flexibility  through  staged  modelling,  Neyman-orthogonal  moments,  cross-fitting  to
reduce bias and overfitting, and without requiring prior assumptions about the relationships among variables38.

We used these two methods because they are based on different causal estimation principles, as noted above,
which enhances the robustness of our results. For both CF and DML, we set SOC as the treatment variable, maize
yield as the outcome variable and the remaining variables (GPR, GTE, TN, AP, AK, pH, N rate, P rate, K rate) as
confounders. These approaches enable quantification of the yield benefits of SOC, expressed as maize yield

Page 9 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

gains per unit increase in SOC. We conducted CF and DML analyses using the R (version 4.3.2) packages ‘grf’33
and ‘DoubleML’39, respectively. Specifically, we used the causal_forest’ function in ‘grf’ with 2,000 causal trees and
other default parameters to run the CF analysis. We implemented DML analysis using RF regressors as learners
(‘lrn’  function)  to  predict  both  the  outcome  variable  (yield)  and  the  treatment  variable  (SOC)  using  default
parameters.  A  partially  linear  regression  model  (‘DoubleMLPLR’  function)  was  then  implemented  within  the  DML
framework  using  these  prespecified  learners  and  fivefold  cross-validation  for  robust  estimation.  Results  from  both
CF and DML are reported as average treatment effects with 95% CIs.

Exploring relationships between SOC and maize yield

We applied RF models to identify the key drivers of maize yield in the farm and field trial datasets, focusing on the
effect and relative importance of SOC. Partial dependence analysis of the RF models was used to visualize the
marginal relationship between SOC and maize yield in both farms and field trials. This method also enabled
identification of the critical SOC threshold for maize yield5. The RF model was developed using the environmental
and managerial variables in the above dataset with the R (version 4.3.2) package randomForest40. Specifically, for
farms, it was constructed with explanatory variables of GPR, GTE, SOC, TN, AP, AK, pH, N rate, P rate and K rate;
for on-farm trials, it was constructed with GPR, GTE, SOC, AP, AK, pH, N rate, P rate and K rate; and for on-station
trials, it was constructed with GPR, GTE, SOC, TN, AP, AK, pH, N rate, P rate and K rate. All three models used
maize yield as the dependent variable (Supplementary Table 3).

A grid-search procedure was used to select the best hyperparameters (the number of trees, ntree, and the number
of variables at each node, mtry) for the RF models based on the lowest cross-validation error (Supplementary Table
4). The optimal values of ntree and mtry were 500 and 5 for farms, 1,600 and 3 for on-farm trials, and 100 and 5
for on-station trials, respectively. We validated the RF models using fivefold cross-validation with stratified random
sampling.  We  evaluated  model  performance  on  the  test  sets  using  the  coefficient  of  determination  (R2)  and  root
mean square error, with the final metrics reported as the mean across all folds. RF models performed well on the
test  sets  for  on-station  trials,  with  R2  values  of  0.62,  compared  with  0.43  for  farms  and  0.42  for  on-farm  trials
(Supplementary  Fig.  13).  Predictor  importance  in  the  RF  models  was  assessed  using  the  percentage  increase  in
mean squared error (%IncMSE). Multiple independent runs with different random seeds confirmed the robustness
of the ranking of the most important predictors (Supplementary Fig. 14). To investigate the potential impact of data
structure  and  distribution  on  RF  models,  we  further  incorporated  spatial/group  information  (coded  as  the  factor
‘site’) as an additional predictor. Whether the site factor was included or not had minor effects on the validated R2
values  and  SOC–yield  relationships  (Supplementary  Fig.  15).  Partial  dependence  analysis  was  conducted  using
the R (version 4.3.2) package pdp41 and subsequently smoothed using a generalized additive model implemented
in the R (version 4.3.2) package mgcv42. The 100-time bootstrap approach was used to estimate the uncertainty
associated with the partial dependence analysis, in accordance with previous research5. The rfPermute package43
was used to estimate the statistical significance of each predictor.

To further understand the SOC–yield relationship, we used SHAP analysis44. For SOC, positive SHAP values
indicate increases in predicted maize yield, whereas negative SHAP values indicate decreases in predicted maize
yield. SHAP values for RF models were computed using the R (version 4.3.2) packages fastshap and shapviz, with
the explain function and 1,000 Monte Carlo repetitions to ensure numerical stability and reliability.

Evaluating the effects of environments and management practices

We  used  a  combined  approach  of  a  sliding-window  analysis  and  partial  correlation  to  evaluate  how  soil,  climate
and management factors modulated the correlation between SOC and maize yield. This analysis is well suited to
investigating the trend of the correlation between two factors in response to the gradient of the third one45,46. It
could  be  applied  to  evaluate  whether  modulating  effects  are  linear  or  abrupt,  which  was  more  refined  and
meaningful  than  the  conventional  grouping  approach.  In  this  study,  we  evaluated  the  modulating  effects  of  GPR,
GTE, pH, clay, N rate, P rate, K rate, no-till, straw return and manure application on the partial correlation between
SOC and maize yield in both farms and field trials. For the continuous variables (GPR, GTE, pH, clay, N rate, P
rate  and  K  rate),  we  first  ordered  each  variable  starting  from  the  lowest  to  the  highest  value,  along  with  all  other

Page 10 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

variables in the dataset. We then selected a window width of 50 steps for both farms and on-station trials and 200
steps  for  on-farm  trials  because  the  on-farm  trials  had  a  larger  observation  size  (n  =  1,030)  than  the  farms  (n  =
192)  and  on-station  trials  (n  =  216).  For  each  window  width,  we  calculated  the  mean  value  of  every  moderating
variable and the partial correlation between SOC and maize yield. For the categorical variables (use or no use of
till, straw return and manure application), we first grouped the datasets according to the categorical variables and
then  calculated  the  partial  correlation  between  SOC  and  maize  yield  separately.  Note  that  this  analysis  was
applied only to the farms and on-station trials because the soil management practices in on-farm trials were the
same:  conventional  tillage  without  straw  return  and  manure  application.  When  calculating  partial  correlations,  we
controlled  all  indicators  except  SOC  to  eliminate  the  influence  of  other  pedoclimatic  and  managerial  factors.  The
response  of  the  SOC–yield  partial  correlation  to  each  factor  was  quantified  using  linear  and/or  nonlinear  models
(Supplementary Table 5). The sliding-window analysis was performed using the R (version 4.3.2) packages zoo47
paired with ppcor48 for partial correlation analyses.

Uncertainty and sensitivity analyses

We  conducted  uncertainty  and  sensitivity  analyses  to  assess  the  robustness  of  SOC  effects  on  maize  yield  by
accounting for data quality and modelling choice. First, model input data may affect our findings. We conducted a
Monte Carlo simulation to quantify uncertainties arising from both data sources and the CF, DML and RF models.
All  model  inputs,  including  climate  variables  (GPR,  GTE),  soil  properties  (SOC,  TN,  AP,  AK,  pH),  N,  P  and  K
application  rates,  and  maize  yield,  were  assumed  to  follow  normal  distributions  for  both  farms  and  field  trials.
Uncertainty  was  characterized  using  the  coefficient  of  variation  (standard  deviation/mean),  with  the  coefficient  of
variation values sourced from previous studies (Supplementary Table 6). We ran 1,000 iterations for all models, and
the  results  are  summarized  as  means  with  95%  prediction  intervals  constructed  using  the  2.5th  and  97.5th
quantiles.  Uncertainty  from  model  misspecification  was  also  quantified  through  variation  in  outcomes  across
different  model  structures  and  parameter  settings.  We  assessed  variability  in  model  predictions  under  different
hyperparameter settings (for example, the number of trees and the number of features). Algorithmic instability was
determined  from  the  variability  in  model  estimates  across  multiple  fits  to  the  same  simulated  data.  Results
accounting  for  all  sources  of  uncertainty  supported  the  conclusion  that  increasing  SOC  enhanced  maize  yield  in
field trials but had no detectable effect on maize yield in farms across Northeast China (Supplementary Fig. 3).
Partitioning  of  the  total  uncertainty  revealed  that  model-related  uncertainty  was  the  dominant  contributor  (~95%)
relative to other sources (Supplementary Fig. 16).

Second, we excluded data points that showed large discrepancies between farms and field trials. This analysis
focused on key variables: SOC, GPR and GTE. Outliers were identified using a cross-group 3

σ

rule: the lower

threshold was set as the maximum of (

μ

(mean value)

−

3

σ

) and the upper threshold as the minimum of (

μ

+ 3

σ

),

calculated separately for farms, on-farm and on-station trials (Supplementary Table 7). After removing outliers for
SOC, GPR and GTE, we re-ran the CF and DML analyses. The results remained consistent with those from the full
dataset (Supplementary Fig. 2), confirming that our causal conclusions are robust to the exclusion of SOC and
climate outliers. Together, our uncertainty and sensitivity analyses demonstrated that the divergent SOC effects on
maize yield between farms and field trials remain robust to variations in data quality and modelling adjustments.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Acknowledgements

We thank all farmers and county government staff who supported the study. We thank all the people who
participated in the survey and soil sampling. We thank S. Wang and X. Cui for their helpful comments on the
methodology in an earlier version of this paper.

Notes

Page 11 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

Supplementary informationThe online version contains supplementary material available at

https://doi.org/10.1038/s43016-026-01401-5

.Publisher’s note Springer Nature remains neutral with regard to

jurisdictional claims in published maps and institutional affiliations.

Bibliography

[1].Nerini, FF. Extending the Sustainable Development Goals to 2050—a road mapNature (2024). Vol. 630. Pg 555-
558.

[2].Nationally

Determined

Contributions

(NDC)

under

the

Paris

Agreement

Synthesis

Report

FCCC/PA/CMA/2024/10 (UNFCCC Secretariat, 2024).

[3].Bossio, DA. The role of soil carbon in natural climate solutionsNat. Sustain. (2020). Vol. 3. Pg 391-398.

[4].Lal, R. Soil organic matter content and crop yieldJ. Soil Water Conserv. (2020). Vol. 75.

[5].Ma, Y. Global crop production increase by soil organic carbonNat. Geosci. (2023). Vol. 16. Pg 1159-1165.

[6].Lal, R. Soil carbon sequestration impacts on global climate change and food securityScience (2004). Vol. 304.
Pg 1623-1627.

[7].Oldfield, EE. Bradford, MA. Wood, SA. Global meta-analysis of the relationship between soil organic matter
and crop yieldsSoil (2019). Vol. 5. Pg 15-32.

[8].Oldfield, EE. Positive associations of soil organic matter and crop yields across a regional network of working
farmsSoil Sci. Soc. Am. J. (2022). Vol. 86. Pg 384-397.

[9].Zhang, W. Closing yield gaps in China by empowering smallholder farmersNature (2016). Vol. 537. Pg 671-
674.

[10].Moinet, GY. Hijbeek, R. van Vuuren, DP. Giller, KE. Carbon for soils, not soils for carbonGlob. Change Biol.
(2023). Vol. 29. Pg 2384-2398.

[11].Zhao, J. For the protection of black soilsNat. Food (2025). Vol. 6. Pg 119-120.

[12].Zhao, Y. Economics- and policy-driven organic carbon input enhancement dominates soil organic carbon
accumulation in Chinese croplandsProc. Natl Acad. Sci. USA (2018). Vol. 115. Pg 4045-4050.

[13].National Bureau of Statistics of China Chinese Statistical Yearbook 2024 (China Statistics Press, 2025).

[14].Oldfield, EE. Wood, SA. Bradford, MA. Direct evidence using a controlled greenhouse study for threshold
effects of soil organic matter on crop growthEcol. Appl. (2020). Vol. 30.

[15].Ordóñez, RA. Insufficient and excessive N fertilizer input reduces maize root mass across soil typesField
Crops Res. (2021). Vol. 267.

[16].Xing, H. Excessive nitrogen application under moderate soil water deficit decreases photosynthesis,
respiration, carbon gain and water use efficiency of maizePlant Physiol. Biochem. (2021). Vol. 166. Pg 1065-1075.

[17].National Soil Survey Office Chinese Soil (China Agricultural Press, 1998).

[18].Jeffery, S. Biochar boosts tropical but not temperate crop yieldsEnviron. Res. Lett. (2017). Vol. 12.

Page 12 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

[19].Guo, N. A quantitative review of the effects of biochar application on the reduction of Cu concentration in plant:
a meta-analysisEnviron. Sci. Pollut. Res. (2024). Vol. 31. Pg 60731-60748.

[20].Guo, JH. Significant acidification in major Chinese croplandsScience (2010). Vol. 327. Pg 1008-1010.

[21].Hu, Z. Nutrient-induced acidification modulates soil biodiversity-function relationshipsNat. Commun. (2024).
Vol. 15.

[22].Daniel, AK. Bradford, MA. Fuller, M. Oldfield, EE. Wood, SA. Soil organic matter protects US maize yields
and lowers crop insurance payouts under droughtEnviron. Res. Lett. (2021). Vol. 16.

[23].Hijbeek, R. Do organic inputs matter—a meta-analysis of additional yield effects for arable crops in
EuropePlant Soil (2017). Vol. 411. Pg 293-303.

[24].Garcia-Vila, M. Gaps and strategies for accurate simulation of waterlogging impacts on crop productivityNat.
Food (2025). Vol. 6. Pg 553-562.

[25].Zhao, X. Crop yields under no-till farming in China: a meta-analysisEur. J. Agron. (2017). Vol. 84. Pg 67-75.

[26].Vonk, WJ. European survey shows poor association between soil organic matter and crop yieldsNutr. Cycl.
Agroecosyst. (2020). Vol. 118. Pg 325-334.

[27].Jiang, R. F. & Cui, J. Y. in Fertilization Technology Highlights (ed. Zhang, F. S.) Ch. 5 (China Agricultural Univ.
Press, 2006).

[28].Qiao, L. Soil quality both increases crop production and improves resilience to climate changeNat. Clim.
Change (2022). Vol. 12. Pg 574-580.

[29].Van Bemmelen, J. Der 851 in Den Colloïdalen Silikaten Gebundenen Kieselsäure, Des Mangans U. S. W. Im
Ack-852 erbodenLandwirthsch. Vers.-Stat. (1890). Vol. 37. Pg 279-290.

[30].Friedrich, S. Schmidt, A. Nonparametric estimation of the coefficient of overlapping—theory and empirical
applicationComput. Stat. Data Anal. (2006). Vol. 50. Pg 1583-1596.

[31].Pastore, M. Overlapping: a R package for estimating overlapping in empirical distributionsJ. Open Source
Softw. (2018). Vol. 3.

[32].Feuerriegel, S. Causal machine learning for predicting treatment outcomesNat. Med. (2024). Vol. 30. Pg 958-
968.

[33].Breiman, L. Random forestsMach. Learn. (2001). Vol. 45. Pg 5-32.

[34].Deines, JM. Wang, S. Lobell, DB. Satellites reveal a small positive yield effect from conservation tillage across
the US Corn BeltEnviron. Res. Lett. (2019). Vol. 14.

[35].Kluger, DM. Owen, AB. Lobell, DB. Combining randomized field experiments with observational satellite data to
assess the benefits of crop rotations on yieldsEnviron. Res. Lett. (2022). Vol. 17.

[36].Altey, S. Wager, S. Estimating treatment effects with causal forests: an applicationObs. Stud. (2019). Vol. 5. Pg
37-51.

[37].Chernozhukov, V. Double/debiased machine learning for treatment and structural parametersEconomet. J.
(2018). Vol. 21. Pg C1-C68.

[38].Jonathan, F., Philipp, B. & Dominik, P. Estimating causal effects with double machine learning—a method
evaluation. Preprint at

https://doi.org/10.48550/arXiv.2403.14385

(2024).

Page 13 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

[39].Bach, P. Kurz, MS. Chernozhukov, V. Spindler, M. Klaassen, S. DoubleML: an object-oriented implementation
of double machine learning in RJ. Stat. Softw. (2024). Vol. 108. Pg 1-56.

[40].Liaw, A. Wiener, M. Classification and regression by random forestR. News (2002). Vol. 2. Pg 18-22.

[41].Greenwell, BM. pdp: an R package for constructing partial dependence plotsR J. (2017). Vol. 9. Pg 421-436.

[42].Wood, S. N. Generalized Additive Models: An Introduction with R 2nd edn (Chapman and Hall/CRC, 2017).

[43].Archer, E. rfPermute: estimate permutation p-values for random forest importance metrics. R Package v 2.5.4
(2025).

[44].Lundberg, SM. From local explanations to global understanding with explainable AI for treesNat. Mach. Intell.
(2020). Vol. 2. Pg 56-67.

[45].Gross, N. Unforeseen plant phenotypic diversity in a dry and grazed worldNature (2024). Vol. 632. Pg 808-814.

[46].Berdugo, M. Gaitán, JJ. Delgado-Baquerizo, M. Crowther, TW. Dakos, V. Prevalence and drivers of abrupt
vegetation shifts in global drylandsProc. Natl Acad. Sci. USA (2022). Vol. 119.

[47].Zeileis, A. Grothendieck, G. Zoo: S3 infrastructure for regular and irregular time seriesJ. Stat. Softw. (2005).
Vol. 14. Pg 1-27.

[48].Kim, S. ppcor: an R package for a fast calculation to semi-partial correlation coefficientsCommun. Stat. Appl.
Methods (2015). Vol. 22. Pg 665-674.

[49].Zhao, S. X. et al. The beneficial relation between soil organic carbon and maize yield in field trials does
not translate to real-world farms in Northeast China. figshare

https://doi.org/10.6084/m9.figshare.32529207

(2026).

[50].Zhao, S. X. The beneficial relation between soil organic carbon and maize yield in field trials does not
translate to real-world farms in Northeast China. figshare

https://doi.org/10.6084/m9.figshare.32539239

(2026).

Classification

Language: ENGLISH

Publication-Type: Magazine

Journal Code: 43016

Subject:  SOIL HEALTH (92%); SOIL CONSERVATION (90%); CORN FARMING (90%); CROP RESEARCH
(90%); AGRICULTURE (89%);  FERTILIZERS (89%); FOOD INDUSTRY (89%);  CARBON CAPTURE &
STORAGE (89%); CONSERVATION (89%);  CARBON SEQUESTRATION (89%); CROP PRODUCTION (78%);
SUSTAINABILITY (78%); FOOD & BEVERAGE (78%); SUSTAINABLE DEVELOPMENT (78%);  EMISSIONS
(73%); SUSTAINABLE DEVELOPMENT GOALS (73%); EVIDENCE (72%); PRODUCTIVITY (65%);
AGRICULTURAL WASTES (78%); SOIL CONSERVATION (90%); SUSTAINABLE DEVELOPMENT GOALS
(73%); CARBON SEQUESTRATION (89%); CARBON CAPTURE & STORAGE (89%); EVIDENCE (72%); CROP
RESEARCH (90%); SOIL HEALTH (92%); SUSTAINABILITY (78%); PRODUCTIVITY (65%); SUSTAINABLE
DEVELOPMENT (78%); EMISSIONS (73%); CONSERVATION (89%)

Page 14 of 14

The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world

farms in Northeast China

Organization: Institute of Carbon Neutrality, Laboratory for Earth Surface Processes, College of Urban and
Environmental Sciences, Sino-France Institute of Earth System Science; Peking University; School of Agriculture
and Food Sustainability; The University of Queensland; Institute of Applied Ecology; Chinese Academy of Sciences;
College of Life Sciences; Northeast Agricultural University; Southwest United Graduate School

Industry: AGRICULTURAL WASTES (78%); AGRICULTURE (89%); SUSTAINABLE DEVELOPMENT GOALS
(73%); FOOD INDUSTRY (89%); FOOD & BEVERAGE (78%); FOOD & BEVERAGE OVERVIEW (78%); CROP
RESEARCH (90%); SOIL HEALTH (92%); CORN FARMING (91%); SUSTAINABLE DEVELOPMENT (78%);
EMISSIONS (73%); CROP PRODUCTION (89%); FERTILIZERS (89%)

Geographic: BEIJING, CHINA (57%); BEIJING, CHINA (57%); NORTHEAST CHINA (92%); NORTHEAST
CHINA (92%); CHINA (94%); CHINA (94%)

Load-Date: August 20, 2026

End of Document

Zhao et al. (2026). The beneficial relation between soil organic carbon and maize yield in field trials does not translate to real-world farms in Northeast China

Page 1

Page 2

Page 3

Page 4

Page 5

Page 6

Page 7

Page 8

Page 9

Page 10

Page 11

Page 12

Page 13

Page 14

Page 15

Page 16

Document Outline