Field Crops Research 297 (2023) 108942

Contents lists available at ScienceDirect

Field Crops Research

journal homepage: www.elsevier.com/locate/fcr

A farmer data-driven approach for prioritization of agricultural research

and development: A case study for intensive crop systems in the
humid tropics
Gonzalo Rizzo a, Fahmuddin Agus b, Siti Fatimah Batubara b, José F. Andrade a, ‡,
Juan I. Rattalino Edreira a, Dwi K.G. Purwantomo c, Rahmat Hanif Anasiru d, Maintang e,
Oswald Marbun e, Rina D. Ningsih b, Syahri b, Baiq S. Ratna e, Via Yulianti e, Nurul Istiqomah b,
Vina Eka Aristya b, Réka Howard f, Kenneth G. Cassman a, Patricio Grassini a, *
a
Department of Agronomy and Horticulture, University of Nebraska-Lincoln, NE, USA
b
Research Organization for Agriculture and Food, National Research and Innovation Agency (BRIN), Cibinong 16911, Indonesia
c
Indonesian Center for Agricultural Land Resources Research and Development, Jalan Tentara Pelajar No. 12, Bogor 16124, Indonesia
d
Indonesia Agricultural Engineering Polytechnic, Serpong, Tangerang 15338, Indonesia
e
Agency for Agricultural Instrument Standardization (BSIP), Jalan Ragunan No.29, Jakarta Selatan 12540, Indonesia
f
Department of Statistics, University of Nebraska-Lincoln, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Context: Intensive rice-maize sequences in Southeast Asia can include up to three crop cycles per year. Indonesia is
Maize the third and fifth largest rice and maize producing country worldwide, and domestic demand for both crops will
Humid tropics increase in the future. Novel, cost-effective and less time-consuming approaches are needed to identify causes of
Rice
yield gap at national level.
Survey data
Yield gap
Objectives: Here, we propose a farmer data-driven approach to prioritize investment in agricultural research and
development (AR&D) programs.
Methods: We collected data on yield, management practices, and socioeconomic variables from 1,147 small­
holders’ fields in intensive rice and maize cropping systems, from 2017 to 2018, across ten provinces in
Indonesia, which include a wide range of landscape positions (upland, lowland, tidal), water regimes (irrigated
and rainfed), and cropping intensities (from single to three cycles per year on the same piece of land). Separate
data were available for each rice and maize cycle included in the annual crop sequence. We used conditional
inference trees, random forest regression, and comparisons among high- versus low-yield fields to identify key
agronomic and socioeconomic factors explaining yield variation.
Results: For a given field and crop species, there was a significant positive correlation between yield in one season
and that in subsequent seasons. In contrast, there was poor correlation between rice and maize yields in cropping
systems including both crops. Socio-economic factors such as years of farming experience and access to extension
services and inputs explain variation in average yield gap across provinces. In turn, agronomic factors such as
nutrient input rates, splits and timing, establishment date, and pest control, explained yield gaps in farmer fields.
Overall, these findings were not consistent with expectations from local researchers about on-farm yield
constraints.
Conclusions: Our study shows that a modest investment to gather farmer survey data, together with robust spatial
frameworks to guide data collection, proper statistical methods to analyze the data, and crop modeling to es­
timate yield potential, can help identify yield constraints for areas representing millions of hectares of rice and
maize.
Significance: Our study provides useful information for guiding investments in AR&D programs at national and
sub-national level for improving crop production by closing current yield gaps.

* Corresponding author.
E-mail address: [email protected] (P. Grassini).
‡
Current address: Cátedra de Cerealicultura, Facultad de Agronomía, Universidad de Buenos Aires, IFEVA-CONICET, 1417, Buenos Aires, Argentina.

https://doi.org/10.1016/j.fcr.2023.108942
Received 23 November 2022; Received in revised form 29 March 2023; Accepted 11 April 2023
Available online 21 April 2023
0378-4290/© 2023 Elsevier B.V. All rights reserved.
G. Rizzo et al. Field Crops Research 297 (2023) 108942

1. Introduction of Southeast Asia through analysis of farmer data or field trials (e.g.,
Dobermann et al., 2004; Boling et al., 2010; Laborte et al., 2012;
Intensive rice-maize crop sequences are common in the humid tro­ Pasuquin et al., 2014; Stuart et al., 2016; Silva et al., 2018; Stuart et al.,
pics where warm temperature and ample water resources permit up to 2018; Silva et al., 2022). However, these studies focused on individual
three crops on the same piece of land each year (Swastika et al., 2004; crops and crop cycles typically at only a few sites and did not have an
Pasuquin et al., 2014; Agustiani et al., 2018; Agus et al., 2019). explicit goal of providing input to prioritization of agricultural research
Indonesia is the main rice and maize producing country in Southeast and development (AR&D) to explore opportunities for increasing
Asia, accounting for 33% and 50% of total rice and maize production in self-sufficiency at national scale in staple food crop production. Like­
this region, respectively (FAOSTAT, 2012–2019; http://www.fao.org/­ wise, previous studies relied on survey data based on farmer-reported
faostat/en/#home). In turn, Indonesia is the fourth most populous yield and input rates “per ha”, which is problematic because farmers
country in the world with ca. 273 million people in 2020 and 310 million typically do not know the size of their fields with accuracy (Fraval et al.,
expected by 2035 (U.N., 2019). Demographic and dietary changes will 2019). Lack of accurate field size measurements leads, especially in
increase national demand for rice and, especially, maize over the next smallholder fields, to extremely large and small yield and input values
three decades (IFPRI, 2017; Arifin et al., 2018; Yuan et al., 2022). and spurious relationships (Gourlay et al., 2017; Desiere and Jolliffe,
Indonesian government policies seek to avoid costly rice and maize 2018; Lobell et al., 2020).
imports by reaching self-sufficiency and exploit opportunities to become A farmer data-driven approach can be used to identify major yield
a net exporter of these crops (Dawe and Timmer, 2012; Buresh et al., constraints for individual crops in specific regions (e.g., Grassini et al.,
2021). But Indonesia is not an isolated case, indeed, interest in food 2015a; Stuart et al., 2016; Rattalino Edreira et al., 2017, 2020). Such an
self-sufficiency (or degree of self-sufficiency that is feasible) at national, approach can help minimize time lost to trial and error, providing a
continental, and international scales has been renewed given recent focus to field studies aiming to identify causes for yield gaps and eval­
events such as the war in Ukraine starting in Feb 2022, extreme weather uate management practices to overcome them, ultimately making AR&D
events, and COVID-19 pandemic (Brankov et al., 2021; Clapp, 2017; investments more efficient (e.g., Andrade et al., 2022a). Three steps are
Kriewald et al., 2019; https://www.nytimes. needed for cost-effective implementation of this data-driven prioritiza­
com/2022/06/10/climate/food-farming-climate-change.html; tion approach at national scale: (1) identify the most representative crop
https://nationalinterest.org/tag/food-security). producing regions within a country and focus data collection to cover
Yield potential (Yp) of a well-adapted cultivar is determined by at­ the most dominant cropping systems and soil types within each region
mospheric carbon dioxide (CO2) concentration, temperature, and solar based on contribution to total national production, (2) use of statistical
radiation in absence of limitations from water, nutrients, weeds, path­ methods to accommodate spatially and temporally explicit data sets
ogens, and insect pests (Evans, 1993). For rainfed crops, water supply obtained from farm- or field-level scale, and (3) interpretation of results
and soil properties influencing crop water balance also determine the based on agronomic principles (i.e., crop genetics, crop ecology, soil
water-limited yield potential (Yw). In the case of intensive cropping science). We hypothesize that application of this data-driven approach
sequences, where more than one crop is grown sequentially on the same identifies the most likely causes of Yg and agronomic inefficiencies (i.e.,
piece of land within a 12-month period, each crop cycle has a different leading to higher costs and increased pollution), which, in turn, provides
Yp (or Yw) due to differences in establishment date and, thus, weather initial research targets for AR&D investments at national and regional
conditions (and stored soil moisture leftover from previous crop in the scale that are supportive of national food security. We are not aware of
case of rainfed crops). Average yields from farmer fields are typically studies combining farmer yield and management data with soil and
well below Yp (or Yw) because of yield-limiting and reducing factors climate databases at national scale to prioritize investments in AR&D
such as nutrient deficiencies, weed competition, and damage from insect programs to improve their impact through focus on the most limiting
and disease pests (van Ittersum and Rabbinge, 1997; Cassman, 1999). factors.
The difference between Yp (or Yw) and farmer yield is referred to as the Here, we propose a farmer data-driven approach that, with a modest
yield gap (Yg). As a reference, a farmer with good access to inputs, investment, can efficiently detect the main yield constraints and serve as
markets, and technical knowledge could reach approximately 70–80% a basis to prioritize investment on AR&D. To test this concept, we used a
of Yp (or Yw) (Cassman et al., 2003). Beyond this level, it would be database that includes farmer-reported data from 1,147 fields across the
difficult to achieve further cost-effective yield gains with existing tech­ ten main rice and maize producing provinces of Indonesia, together with
nologies. Previous studies have reported a moderate Yg for rice in information for all crops involved within each crop sequence. Actual
Indonesia, with farmer yields averaging 63% of Yp for irrigated rice and yield was derived from the farmer-reported data, Yp (or Yw) was
53% of Yw for lowland rainfed rice (Agus et al., 2019). The reported Yg retrieved from Agus et al. (2019), and the Yg was determined by dif­
for maize was much larger, with farmer yields averaging 44% and 42% ference. Finally, we compared our list of yield constraints with causes for
of Yp (irrigated) and Yw (rainfed), respectively. Hence, considering the Yg suggested by local scientists and discussed implications of our find­
large area sown with these crops (ca. 11 and 5 M ha for rice and maize, ings for prioritization of investments on AR&D and inform policy.
respectively), a substantial increase in aggregate national production
could be expected if causes for Yg are identified and overcome via 2. Materials and methods
improved agronomic management and cultivars.
Perceptions of scientists alone, without field validation, are not 2.1. Site selection
sufficient to identify yield constraints and prioritize research (Cassman,
2001; Hunter et al., 2017; Cassman and Grassini, 2020). At the same We focused on the top 10 Indonesian provinces accounting for the
time, identifying yield constraints through traditional on-farm trials is largest share of rice and maize harvested area: North Sumatra (NSU),
labor-intensive, expensive, and time-consuming—often requiring South Sumatra (SSU), Lampung (LAM), West Java (WJA), Central Java
several years across sufficient locations to capture the range of envi­ (CJA), East Java (EJA), South Kalimantan (SKA), South Sulawesi (SUL),
ronments across time and space in the targeted region. When farmer Gorontalo (GOR), and West Nusa Tenggara (WNT) (Fig. 1). All together,
data on yield and management practices are available and properly these provinces represent 73% and 80% of the respective national rice
contextualized in terms of climate and soil, analysis of such data can and maize harvested area (https://www.bps.go.id/indicator/53/21/1/
help identify factors responsible for Yg and quantify their impact (Cock harvested-area.html). We targeted both rice and maize in six of the ten
et al., 2011; Grassini et al., 2011; Rattalino Edreira et al., 2017; Andrade provinces (NSU, LAM, CJA, EJA, SUL, and WNT) because the area
et al., 2022a). Several studies have attempted to identify causes of Yg in planted with both crops is large (>100,000 ha). In three of these six
intensive cereal systems in Indonesia and elsewhere in the humid tropics provinces, lowland rice is grown sequentially with maize (CJA, EJA, and

2
G. Rizzo et al. Field Crops Research 297 (2023) 108942

WNT) while in the other three provinces (NSU, LAM, and SUL), rice and and crop sequence (as determined by crop intensity and temporal
maize are grown in different ecosystem (maize is grown in uplands while arrangement). Because rice and maize are grown sequentially within the
rice is grown in lowland and tidal ecosystems) (Fig. 2). In the remaining same cropping system in three buffers (CJA, EJA, and WNT), the number
four provinces, we focused only on rice (SSU, WJA, and SKA) or maize of selected cropping systems (n = 13) was smaller than that of
(GOR). Hence, a total of 16 crop-province combinations were studied. crop-province combinations (n = 16). Selected cropping systems
Details on main cropping systems in Indonesia, and associated climate included irrigated rice, rainfed lowland rice (including tidal systems),
and soils, are available elsewhere (Swastika et al., 2004; Agustiani et al., irrigated maize, rainfed lowland, and upland maize crops. For data
2018; Agus et al., 2019). collection purposes, we focused on one regency within each province.
Agus et al. (2019) estimated rice and maize Yw (or Yp) and Yg for a Regency is the second-level administrative unit of Indonesia, beneath
number of sites across Indonesia. These authors followed the protocols the provincial level, with an average size of 360,000 ha. Each regency
of the Global Yield Gap Atlas (Grassini et al., 2015b; van Bussel et al., was selected based on the following criteria: (i) accounts for a large
2015; www.yieldgap.org) to select most important rice and maize pro­ portion of total harvested area within each buffer, (ii) includes the
ducing areas across the Indonesian archipelago in terms of contribution dominant crop sequence in the buffer, (iii) is located within the target
to national food production. Sites were selected by (i) a climate zone buffer, and (iv) is representative of dominant agricultural soil in the
(CZs) scheme that delineates geographical areas based on spatial vari­ buffer as determined from high-resolution soil maps provided by experts
ation in growing degree days, temperature seasonality, and aridity index from the Indonesian Center for Agricultural Land Resources Research
(van Wart et al., 2013), and (ii) harvested area distribution for each and Development (https://www.litbang.pertanian.go.id/produk/67/).
crop-water regime combination around year 2010 (IFPRI, 2019; The five villages selected for each of the 13 cropping systems were
SPAM2010; http://www.mapspam.info). First, CZs that accounted for located within a 25-km radius area to facilitate the survey process. We
> 5% of total national harvested area for each crop-water regime were avoided villages that were not representative of the dominant cropping
selected. Subsequently, within each selected CZ, polygons of 100-km sequence and/or located close to large urban centers.
radius (hereafter referred to as ‘buffers’) were created around existing
weather stations. Buffers were clipped by CZ’s borders to minimize 2.2. Data collection and quality control
weather variation within buffers. In the present study, we selected the
buffer with largest harvested area of the target crops for each of the 16 Personnel from the Assessment Institute of Agricultural Technology
crop-province combinations. Eleven of the sixteen were situated entirely (AIAT) and local extension workers were trained during in-person
within a unique CZ. In cases in which there were two or more buffers (for workshops to ensure consistency on protocols for data collection. Data
a given crop) located within the same CZ, we decided to keep them all were collected manually via individual personal interviews with
for our analysis to account for differences in crop sequences and soil type farmers. We selected 20 farmers in each of the five villages within each
among buffers and to achieve a better representation of CZs with large province (100 surveyed farmers per cropping system in each province).
crop harvested area. Selected buffers were located in CZs that account Within each province, farmers with diverse household income, age,
for 87% and 79% of respective rice and maize national harvested area in education, etc. were selected to properly account for variation in socio-
Indonesia, with buffers including, on average, ca. one third of the crop economic background. We did not interview farmers with fields (i)
area in the CZ where they were located. Hence, our site selection can be including crops other than maize or rice, (ii) planted with extremely rare
considered representative of the main rice and maize producing areas in rice cultivars, open pollinated varieties of maize, or intercropping, (iii)
the country. planted with multiple varieties or with portions planted at different
In each buffer, we focused on the main rice and/or maize cropping dates (more than a 7-day difference), and (iv) used for experimental or
systems (Fig. 2). Cropping system refers to a combination of ecosystem demonstration trials, (v) following outdated management practices (e.g.,
(upland, lowland, tidal; IRRI, 1984), water regime (irrigated, rainfed), oxen-pulled plough). We collected data from one field per farmer; the

Fig. 1. Selected provinces (light red) and re­

gencies (dark red) for assessing causes of yield
gaps in rice and maize systems in Indonesia:
North Sumatra (NSU), South Sumatra (SSU),
Lampung (LAM), West Java (WJA), Central
Java (CJA), East Java (EJA), South Kalimantan
(SKA), South Sulawesi (SUL), Gorontalo (GOR),
and West Nusa Tenggara (WNT). Circles show
location of weather stations used for simulating
yield potential. Inset shows rice (green) and
maize area (red) as reported by SPAM2010
(IFPRI, 2019).

3
G. Rizzo et al. Field Crops Research 297 (2023) 108942

Fig. 2. Selected 13 rice-maize cropping systems in each province. Color indicates water regime: irrigated (green), and rainfed (yellow). In some double-rice crop
systems, an additional third rice crop was planted in some fields (dashed boxes). See Fig. 1 for full province names and geographical location.

largest field was chosen for those farmers managing more than one field. maize yields, considering the effect of farm within village, and village as
The survey included a structured questionnaire on farmer socio- random effect, as defined by the following model:
economic background and yield and agronomic practices related to
yijkl = μ + Si + Vj + F(V)kj + eijkl (1)
the target field (Fig. S1). Farmers were asked to provide the amount of
harvested grain, applied inputs, and agronomic practices for each crop
where yijkl is the lth observation for the ith season (S) in the kth farm (F)
grown in the field over the past 12-month period. To avoid errors due to
within the jth village (V), μ represents the overall mean, and e represents
uncertainty in field size, farmers were asked to report harvested grain ( ) ( )
and applied inputs per field rather than per hectare. In the case of har­ the error, assuming: Vj ∼ N 0, σ 2v ; F(V)kj ∼ N 0, σ 2FV ; and
( )
vested grain, we also requested them to indicate whether it was esti­ eijkl ∼ N 0, σ2e . A linear mixed-effect model was fitted for each crop-
mated before or after air drying. For the latter, we assumed standard province separately using lme4 r-package (Bates et al., 2015). We
grain moisture content of 130 and 155 g H2O kg− 1 for rice and maize, could not evaluate the season effects for those cropping systems with a
respectively). For those cases in which harvested grain was estimated single rice or maize cycle per year (maize in CJA, EJA, and WNT, and
before air drying, we requested local extension workers to provide the rice in SKA).
average grain moisture content for each crop cycle at harvest so that we To determine the size of Yg at each of our sites, we retrieved Yw
can adjust the farmer-reported harvested grain to 130 and 155 g H2O (rainfed crops) or Yp (irrigated crops) values reported in a previous
kg− 1 for rice and maize, respectively. In the case of establishment date, study (Agus et al., 2019). In this previous study, rice and maize Yw and
farmers reported transplanting date for rice (except for SSU where rice is Yp were estimated using well-validated crop simulation models based on
direct seeded) and sowing date for maize. measured weather data and dominant soil types for the three to four
After the interview, surveyors and farmers walked around the field to most important crop sequences at each site. In our study, the Yg was
record GPS coordinates. Then each field was mapped in Google Earth to estimated as the difference between the simulated Yw or Yp reported by
determine its size. Field size was later used to calculate yield and input Agus et al. (2019) and the average farmer yield estimated from our
rates per hectare. Surveyed field data were entered into a digital data­ survey data, separately for each site-season combination.
base and screened to detect extreme values or incomplete entries. Fields At issue is how consistent farmer yield was for a given field across
without reported yield or other key variables such as nutrient fertilizer crop seasons, which provides insight about whether the Yg is persistent
application or establishment date were excluded from the analysis. We over time, or it changes erratically from season to season (Lobell et al.,
also excluded fields for which estimated yields and inputs rates exceeded 2010; Farmaha et al., 2016). In addition, for cropping systems that
reasonable thresholds as well as those for which it was not possible to include rice and maize, assessing the relationship between rice and
retrieve field size. After quality control, the database contained data maize yields for individual fields can help to understand whether high-
from a total of 685 fields sown with rice and 462 fields sown with maize (low-) yielding rice fields are also those achieving high (low) maize
(76% of total surveyed fields). Average rice and maize yield estimated yields and vice versa. To do so, we performed rank correlation analysis to
from the survey data, separately for each province, were in agreement evaluate the relationship between rice (NSU, SSU, LAM, WJA, CJA, EJA,
with those reported in official statistics in the same year (https://www. WNT, SUL) or maize yield ranks (NSU, LAM, SUL, GOR) across crop
bps.go.id/indicator/53/22/1/productivity.html), with an average ab­ seasons and assess the association between maize and rice yields for
solute yield difference of 3% (rice) and 1% (maize). This indicates that those cropping systems including both crops (CJA, EJA, WNT). The
our selected regencies and villages were highly representative of the Kendall’s tau method was used to assess the relationship between ranks
target provinces. using cor.test R package (Kendall, 1938). For each province, Kendall’s
tau was assessed separately, and we reported here the average Kendall’s
tau across cropping systems.
2.3. Assessing sources of yield variation among and within cropping
systems
2.4. Identification of agronomic and socio-economic causes of yield gaps
For each cropping system, analysis of variance (ANOVA) was per­
formed to evaluate the effect of season (dry, wet, late wet) on rice and Socio-economic variables included in our analysis were farmer’s age,

4
G. Rizzo et al. Field Crops Research 297 (2023) 108942

years of school education, years of farming experience, field size, land included season (wet, late wet, and dry) as an independent variable to
ownership (as a percentage of total cultivated land), farm income (as a let the model assess the effect of season and season x management on
percentage of household income), and access to extension services, rice and maize yields. The effect of establishment date was analyzed
labor, and inputs. We used linear regression analysis as a first step to independently from that of crop season. To do so, for each cropping
investigate how variation in socio-economic background across sites system, we computed the difference between the specific establishment
influence the magnitude of the Yg. Subsequently, we assessed the in­ date reported for a given field-season combination and the average
fluence of socio-economic variables at explaining yield variation within establishment date for that season within its province, and we used that
each crop-province combination by comparison of high- (HY) versus difference to evaluate the influence of establishment date on yield
low-yielding (LY) fields. To do so, we first created HY and LY field variation. In the case of nutrient, we considered the total amount of
classes based on the upper and lower terciles of the distribution of element nutrient (N, P, and K) by summing the amount of each nutrient
average field yields for each crop-province combination, respectively. coming from each applied fertilizer and from manure. The nutrient
Subsequently, differences in socio-economic variables between HY and contribution from manure was estimated based on type of organic ma­
LY fields were evaluated for statistical significance using t-tests or, in the terial applied and its applied amount. Nutrient concentrations from re­
case of categorical variables, Chi-square (χ2) tests. ported organic materials were retrieved from Dobermann and Fairhurst
Following Mourtzinis et al. (2018) and Tenorio et al. (2021), we used (2000). The average nutrient contribution from manure was about
a conditional inference tree analysis to identify the agronomic causes of 1.2 kg N ha− 1, 1 kg P ha− 1 and 1.5 kg K ha− 1, being similar for rice and
Yg for rice and maize separately for each of the 16 crop-province com­ maize. Across provinces, we did not find strong correlations between N,
binations, and for the pooled data across Indonesia, separately for each P and K for rice (Pearson’s r range: 0.20–0.36) and maize (Pearson’s r
crop. The average number of observations was 140 (rice) and 96 (maize) range: 0.50–0.67) as most of the N and P were supplied via straight
for analysis of individual crop-province combinations and 1446 (rice) fertilizers rather than blends.
and 670 (maize) for the analysis based on the pooled data. In irrigated We used random forest regression analysis to validate the results of
and rainfed systems with relatively low yield variability, as it is the case our conditional inference tree analysis for the 16 crop-province combi­
of the humid tropics, this number of observations seem sufficient for the nations as well as for the pooled data. Random forest regression analysis
analysis (see Mourtzinis et al., 2018). Conditional inference tree analysis was performed using randomForest package in R (Liaw and Wiener,
was performed using the partykit package in R (Hothorn and Zeileis, 2002) to quantify the importance of each independent variable, which
2015). Briefly, the algorithm tests the null hypothesis of independence was based on how much the mean accuracy of the prediction decreases
between the response variable (i.e., yield) and any of the input variables when a variable is excluded. To normalize values among crop-province
(i.e., management and field variables). The algorithm selects the input combinations, and between crops in the case of the pooled data analysis,
variable with strongest association, measured by a p-value, with the relative variable importance was determined by setting the variable
response variable. Then, a binary split is implemented in the selected with the highest importance to 100 and calculating individual values as
input variable (node) and all steps are recursively repeated. The termi­ a percentage of the most important variable.
nal node accounts for the final subset of fields. As a result, this procedure
outputs a graph that resembles a tree. The sizes of intermediate and
terminal nodes are defined according to pre-specified criteria. In this 2.5. Comparison with expectations from local scientists on yield
analysis, the criterion for the independence test was based on univariate constraints
p-values (alpha = 0.05). To ensure adequate power, besides the p-value,
we made sure that each intermediate node account for a minimum of To synthesize information provided by conditional inference tree
25% of total observations, and a terminal node should contain a mini­ analysis for the 16 crop-province combinations, we computed the
mum of 10 observations in the case of the analyses for each appearance frequency of each agronomic factor (Eq. (1)). To do so, we
crop-province combination, and 50 observations for the analysis based grouped variables into five categories: nutrients (i.e., amount, splits, and
on the pooled data. All these criteria must be met at every step of the timing), pest control (i.e., weeding, fungicide, insecticide, mollusqui­
algorithm so that a variable can qualify for a split. To avoid overfitting cide, nematicide, and rodenticide), variety class (i.e., year of release,
and enhance interpretability, the maximum tree depth was set to 10 source), establishment (i.e., tillage method, establishment date, and
nodes. Explanatory power of the conditional inference tree was calcu­ plant density), and damage (i.e., rats, water stress and/or excess)
lated with the coefficient of determination (r2) and root mean square (Table S1). Following this approach, we estimated the frequency of
error (RMSE). The resulting r2 from our conditional inference trees appearance of each category across the 16 crop-province combinations
ranged from 0.20 to 0.40, which is reasonable for this type of analysis as follows:
and consistent with those reported by previous studies following a ∑N
Ij
similar approach (see Mourtzinis et al., 2018; Lollato et al., 2019; Fj = i=1 × 100 (2)
N
Munaro et al., 2020).
In our conditional inference tree analysis, we included all agronomic where Fj is the frequency of appearance of the j category among all
variables collected via our farmer survey as independent variables conditional inference trees (expressed in %), Ij is equal to one for the ith
(Table S1). Our dependent variable was farmer yield (in Mg ha− 1) for the conditional inference tree if at least one variable included in the jth
analysis of individual crop-province combinations, and relative yield category is present, and zero if none of the variables of the jth category
(farmer yield as percentage of % of Yp and Yw for irrigated and rainfed appears in the ith conditional inference tree, and N is the total number of
crops, respectively) for the analysis based on the pooled data for each conditional inference trees for a given crop.
crop, in order to account for differences in biophysical background and We conducted an additional survey to determine how results from
crop sequences across provinces. Because of the large number of rice and our quantitative yield-gap analysis compare with perceptions of yield
maize cultivars reported by farmers (n = 242 and 154, respectively), constraints held by local, regional, and national experts about rice yield
and their rapid turnover over time, we did not attempt to quantify the constraints. Experts included agronomists, breeders, pathologists,
influence of specific cultivars on yield. Instead, we worked with extension personnel, and senior researchers from Universities and Na­
personnel from the Indonesian Rice Research Institute (BB Padi) and the tional Agricultural Research Institutions. This survey was responded by
Indonesian Center for Cereal Crop Research (ICCCR) to categorize rice 204 experts working on research (n = 143) and extension programs
and maize cultivars into four groups according to their year of com­ (n = 61), all of them located within the 10 selected provinces for rice.
mercial release (before and after year 2000) and source (commercial or Each expert was asked to classify each management practice in three
local breeding). For our conditional inference tree analysis, we also categories in relation to their expected impact at explaining the current

5
G. Rizzo et al. Field Crops Research 297 (2023) 108942

Yg in rice: very important, somewhat important, or less important observed in annual double-crop maize systems, which indicates that the
(Fig. S1). For comparison against results from our statistical analysis, we causes for Yg are persistent over time and space. Rice fields exhibiting
grouped the variables into the same five categories (i.e., nutrients, abrupt changes in rank from one season to another were associated with
establishment, variety, pest control, and damage) and, for each cate­ episodes of rat damage and/or water deficit, while for maize fields large
gory, we determined the frequency of experts considering each of them change in ranks was associated with inconsistency in sowing dates be­
as “very important” for explaining Yg. tween seasons. In contrast, correlation between rice and maize field
ranking was poor (average τ = 0.29) in cropping systems that included
3. Results both crops within the sequence (EJA, CJA, and WNT), indicating that the
fields with ‘highest’ rice yields are not necessarily those where the
3.1. Yield and yield gaps across provinces and crop cycles ‘highest’ maize yields are obtained and vice versa. Finally, the low in­
fluence of village (<17% of total variance) suggests that results are not
Average rice yield in farmer’s fields ranged from 2.1 to 7.0 Mg crop− 1 influenced by village selection and can be extrapolated with confidence
ha− 1 across provinces, with lowest yield in rainfed tidal systems in SSU to other fields that follow the same cropping systems within our target
and SKA, and highest yields corresponding to irrigated systems in WJA provinces (Table 1).
(Fig. 3). There was a significant season effect in most provinces
(Table 1). For those systems including more than one rice cycle (all but
SKA), rice yields were highest and lowest in the dry and late wet season, 3.2. Socioeconomic and agronomic causes of yield variation across
respectively, with intermediate yields in the early wet season (Fig. 3). An farmer fields
exception was LAM with low yield in the dry season because of insuf­
ficient irrigation water supply. In the case of maize, average yield ranged Average measured field size was 0.4 ha for rice and 0.6 ha for maize,
from 3.6 to 7.3 Mg crop− 1 ha− 1, with lowest yield in upland rainfed being smallest in traditional rice and maize producing areas in Java and
systems in the late wet season, and highest yields corresponding to largest in new agricultural areas such as tidal rice in SSA and SKA and
lowland irrigated maize in EJA and WNT. Maize yield was lower in late upland maize in GOR and SUL (Fig. 5). Most farmers (ca. two thirds)
wet versus early wet season in upland maize systems at all sites but NSU. owned the land they managed, except for rice farmers in CJA, and rice
Variation in simulated average Yp (or Yw) across sites and seasons and upland maize farmers in SUL who, on average, owned 37–60% of
followed that observed for average farmer yield (p = 0.05), although their farmed area and the remainder rented land. There was no differ­
temporal variation in Yp was less marked (Fig. 3). The Yg was notably ence in average farmer age (mean: 50 years) and years of education
larger for maize than rice, representing 43% and 52% of the Yp (or Yw), (mean: 7.4 y) between rice and maize farmers (p > 0.25) (Table S3). In
respectively. Regions with long agriculture history and small climatic contrast, rice farmers (either in continuous rice or rice-maize sequences)
risk (as determined from the variation in Yp or Yw) had the smallest Yg, had more years of experience compared with those growing upland
as it was for irrigated rice in WJA and irrigated maize in EJA. In contrast, maize (26 and 20 y, respectively; p = 0.01).
Yg were largest in relatively new and risky production environments Proportion of family income derived from rice and maize production
where crops are often exposed to episodes of water stress and/or excess, was similar between crops and provinces, ranging from ca. 70–90%
such as tidal rice systems in SKA and SSU and upland rainfed maize across crop-province combinations (Fig. 5). Likewise, there was a posi­
systems in Sumatra and Sulawesi islands. tive relationship between average field size and proportion of family
For both crops, there was a significant farm-within-village effect on income from farming among provinces (r2 = 0.27; p = 0.04), suggesting
yield across the 12 crop-province combinations that included two or lower potential to earn adequate income from farming alone in prov­
three rice or maize crops within the crop sequence (Table 1). Fields with inces where the average farm size is smaller. More than two thirds of
highest (or lowest) yields in one season also outperformed (or under­ farmers indicated good access to inputs, labor, and extension services,
performed) other fields in subsequent season(s). These results are except for rice farmers in SSU and SKA and maize farmers in LAM, SUL,
consistent with the Kendall’s tau correlation analysis which found that and GOR, which might be one of the reasons for their comparably low
yield ranks were positively correlated for both rice and maize (average input use and/or large Yg (Fig. 3). Indeed, in rice, there was a correlation
τ = 0.63 and 0.60, respectively) in those systems with more than one between average Yg (expressed as % of Yp or Yw) and differences in
cycle of these crops per year (Fig. 4). In other words, best rice fields in access to labor and inputs across provinces (r > 0.40; p < 0.05), while
one season were also best rice fields in other seasons, and the same was none of these correlations were statistically significant in maize. Dif­
ferences in socio-economic factors between HY and LY fields were

Fig. 3. Farmer-reported rice and maize yields

for each crop season across provinces. Boxes
delimit first and third quartiles, with solid line
and cross inside box representing median and
mean, respectively. Whiskers delimited 10th
and 90th percentiles, with upper and lower dots
representing 5 and 95th percentiles. Red dots
show the simulated yield potential (Yp) for
irrigated crops or water-limited yield potential
(Yw) for rainfed crops for each site as reported
by Agus et al. (2019), while the red error bars
refer to its standard error. Average yield gap
(expressed as % of Yp or Yw) is also shown for
each crop-province combination. See Fig. 1 for
full province names.

6
G. Rizzo et al. Field Crops Research 297 (2023) 108942

inconsistent across cropping systems and provinces and cannot be

within village (F in V) and village (V) as random effects. Degrees of freedom (df) and F-test values (F) are shown for the fixed effects, while percentage of the total random effect variance (%Var) due to F in V, V and residual
Summary of the analysis of variance of the mixed-effects model estimated to assess the effect of season on rice and maize yield across 16 crop-province combinations in Indonesia. The mixed-effects model included farm

F
generalized at the country level (Tables S4 and S5).

-
Average rice hill density was 20 hills m− 2, with lower plant densities

%Var

%Var
WNT

n.c.
84

12

99
df

df
in WJA, LAM and SKA compared with the other provinces (averages: 15

1
-

-
hills versus 23 hills m− 2). Maize average seeding rate was 10 seeds m− 2,
typically arranged in rows 65-cm apart in hills separated within row by

< 1 * **
30 cm, with 1–2 maize seeds per hill. More than one plant per hill were
F

F
reported in ca. 60% of maize fields (n = 336) across all provinces. In the

-
case of CJA, EJA and SUL, nearly all maize fields reported more than one
%Var

%Var
SKA

n.c.
plant per hill. Average N fertilizer rate was 140 and 176 kg N ha− 1 for
57

94
df

df
6

-
rice and maize, respectively, with lowest rates in SKA and SSU (rice) and
NSU and SUL (maize) and largest rates in Java and WNT (Fig. 5). Similar

<1
F

F
patterns across provinces were observed for P and K fertilizer rates. In
-

relation to nutrient fertilizer management, most farmers split N appli­

%Var

%Var
GOR

48

76
11
12
cation in two, typically occurring one and four weeks after transplanting
df

df
-

for rice and two and five weeks after sowing for maize. In contrast, most
P and K fertilizer was applied as a basal application around establish­
< 1 * **

ment date. Details on agronomic management can be found in Supple­

<1

mentary Information Section 1.

F

Asterisks indicate statistical significance at * p < 0.05, * * p < 0.01, and * ** p < 0.001. n.c.: season effect not calculated because a single crop was grown.
Conditional inference tree analysis for pooled data for rice and maize
%Var

%Var
SUL

were selected to illustrate effects of specific factors and interactions

65

90

45

74

26
df

df
4
6

among them ( Fig. 6). In both crops, nutrient-related factors were

identified as the most important management practices explaining yield
153 * **

< 1 * **

gap variation in Indonesia. In the case of rice, the four variables selected
by the conditional inference tree were N and K fertilizer rate, timing of
F

the second urea split (FTS2) and foliar insecticide application. Fields
%Var

%Var
LAM

receiving larger N amounts and higher number of in-season foliar

70

32

66

64

63

28
df

df
2

insecticide applications had, on average, a relative yield of 82% of Yp (or

Yw). Likewise, yield was higher with a second urea split no later than
294 * **

five weeks after transplanting and with larger application of K fertilizer.

In the case of maize, the three variables identified as the most important
F

F
-

in our analysis were N and K fertilizer rate and number of urea splits.
%Var

%Var

Results from the conditional inference tree analysis for all crop-
SSU

63

36
17
47
df

df
-

provinces combinations are shown in the Supplementary Material

(Tables S7-S8).
<1

<1

Results derived from the random forest analysis were consistent with
F

those from the conditional inference trees (Figs. 6 and 7). In the case of
%Var

%Var
NSU

rice, the four variables identified by the conditional inference tree

84

94

59

80

20
df

df
0
6

analysis were within the top ten most important variables identified by
the random forest analysis. However, the conditional inference tree did
< 1 * **

not capture the influence of hill density on yield, which was found to be
the second most important variable as determined by random forest
F

F
-

analysis. In the case of maize, N and K fertilizer were identified as the

%Var

%Var

most important variables by both types of analysis, except for urea split,
EJA

n.c.
85

60

40

96
df

df
0

which was identified as an important variable by the conditional infer­

-

ence tree but did not show up within the top ten most important vari­
are shown for the random effects. See Fig. 1 for full province names.

199 * **

ables as per the random forest analysis.

Across provinces, conditional inference tree analysis using agro­
F

F
-

nomic factors explained 16–66% (rice) and 12–42% (maize) of field-to-

%Var

%Var

field yield variation (Tables S7 and S8). Factors explaining yield varia­
CJA

100
n.c.
85

60
14
26
df

df

tion changed across provinces and crops. In the case of rice, variables
0
-

related to nutrient input rates and management were the most important
factors explaining yield variation in five out of nine provinces. For
< 1 * **

example, N rate was the most important variable explaining yield

F

variation in WJA and SKA, K rate explained most of the variation in NSU
-

and WNT, while timing of the second fertilizer application was the main
%Var

%Var
WJA

69

93

factor in SUL. Season was the main factor in CJA, SSU and LAM, and
df

df
0
7

variety was the most important factor in EJA. Similarly, in the case of
maize, five out of seven provinces showed nutrient rates as the most
Random effect

Random effect

important factors explaining yield variation (N in CJA, SUL, GOR, and

Fixed effect

Fixed effect
Residual

Residual

WNT, and P in LAM). In other cases, establishment date and time of

Season

Season
F in V

F in V

weed control were the most important factors, as it was the case in EJA
V

and NSU, respectively. In maize, conditional inference trees with highest

explanatory power (r2 > 0.40) were those from provinces receiving
Table 1

Maize

lower nutrient fertilizer rates (SUL and GOR) (Table S8). Not surpris­
Rice

ingly, fertilizer rate and timing were identified as the major explanatory

7
G. Rizzo et al. Field Crops Research 297 (2023) 108942

Fig. 4. Kendall’s tau rank correlation analysis between crop cycles of rice (a) and maize crops (b) and between rice and maize fields (c). Kendall’s tau was estimated
separately for each province. Average Kendall’s tau is shown in each panel (p-value < 0.01).

factors for yield variation in these provinces. causes of Yg for major producing areas within a country. This approach
Considering the group of fields following the combination of man­ can help evaluate current AR&D priorities and inform policy options
agement practices leading to highest yields in each crop-province aiming to increase national self-sufficiency in staple food crop produc­
combination, associated yields were, on average, 41% and 38% higher tion. In our case-study on cereal-based systems in Indonesia, we found
than average rice and maize yields, respectively (Tables S7 and S8). In that Yg were larger for maize compared to rice, and Yg was larger in new
turn, these yields represented respective 77% and 63% of current Yp (or agricultural areas relative to the traditional irrigated crop producing
Yw) for rice and maize. Thus, in most provinces, there were rice farmers region in Java, which is consistent with previous reports (Agus et al.,
who have apparently closed most of the exploitable Yg, assuming 80% of 2019; Andrade et al., 2022b). Our study advances previous knowledge
Yp (or Yw) to be a reasonable yield goal. In the case of maize, high- on Yg in the humid tropics of Indonesia by showing that the field yield
yielding fields were still substantially below their potential, except for rank for the same crop remains relatively stable across seasons but that
those in EJA. there is little correlation between yield ranks for rice and maize in
cropping systems including both crops. We attribute this pattern to the
3.3. Results from analysis of farmer data versus perceptions from local longer history of rice versus maize cultivation and associated farmer
experts experience and skills in growing these crops. For example, 60% of the
maize fields (and nearly all in Java) have more than one plant per hill,
When aggregated at national level, our analysis revealed that fer­ which is not a recommended practice for maize (Rossini et al., 2011).
tilizer rates and management (i.e., time of application) explained the Poor soil physical properties for maize grown after puddled rice on clay
largest portion of total yield variation (Fig. 8a). In contrast, pest control soils is also a possible factor explaining the weak correlation between
appeared in 33% and 14% of the conditional inference trees for rice and rice and maize yields (So and Ringrose-Voase, 2000). A reason for
maize, respectively, while the influence of variety and damage from pest consistent ranks in the case of rice could be attributable to the field
and water problems (i.e., excess and/or deficit) was apparent in < 10% having good soil, favorable climate or irrigation system, a competent
of crop-province combinations. We recognize that influence of damage farmer, or a combination of these three factors (Farmaha et al., 2016).
from pest and water problems may be underestimated since we have Our conditional inference tree analysis identified nutrients as the
discarded fields with extreme damage events from our analysis. Estab­ main factor explaining yield variation and, therefore, Yg in rice and
lishment practices, including tillage, plant arrangement, density, and maize crops in Indonesia. This finding was consistent with previous
date of establishment, was second most important group of variables studies showing nutrient management as a key factor influencing on-
explaining yield variation for maize and third most important factor for farm yields (Boling et al., 2004; Dobermann et al., 1998, 2004; Pasu­
rice. quin et al., 2014; Wade et al., 1999). It is noteworthy that the thresholds
Comparison of our findings for rice against perceptions from local for N input rates found in our conditional inference trees were quite
experts about causes for Yg revealed two important discrepancies high. For example, the conditional inference trees for rice and maize in
(Fig. 8b). First, the importance of nutrients as a yield constraint seems to CJA and maize in EJA differentiate among fields applying N fertilizer
be underestimated, as only half of the experts rated this variable as a rates above and below ca. 230 kg N ha− 1 (Tables S7 and S8). This rate of
“very important” factor explaining field-to-field yield variation. Second, N input exceeds N uptake by the rice crop by about a 115%, indicating
ca. 90% of experts rated variety as a “very important” variable, but our that, on average, there is a large N surplus at risk of loss (Buresh et al.,
analysis of farmer data shows that using modern or old varieties has a 2010). This finding points to a low efficiency in the use of N fertilizer due
relatively small impact on yield compared with other agronomic prac­ to poor nutrient management as it has been highlighted by others in
tices. Finally, ca. half of the experts said that establishment was an previous reports (Dobermann et al., 2002; Buresh, 2010; Silva et al.,
important factor influencing rice yields, while our analysis shows this 2018; Silva et al., 2022). Indeed, partial factor productivity (kg grain
factor to be important for maize but less relevant for rice. yield/kg N fertilizer input) for N in Indonesia is low, and N balance
large, compared with that in other rice systems in the region and the
4. Discussion world (Yuan et al., 2021), highlighting an opportunity to increase both
yield and input-use efficiency through better use of the N fertilizer in­
Our study shows that a modest investment to collect and analyze puts (Silva et al., 2022). Similarly, we found fields receiving a high
farmer data, guided by a robust spatial framework based on a CZ scheme number of insecticide applications (>9), which is consistent with the
and appropriate statistical methods, allows identification of major higher use of pesticides in Indonesia in relation to other countries with

8
G. Rizzo et al. Field Crops Research 297 (2023) 108942

Fig. 5. Socio-economic and agronomic vari­

ables collected through farmer interview for
rice (left) and maize (right) fields across 10
provinces. Boxes delimit the first and third
quartiles, with solid line and cross inside the
box representing the median and mean. Whis­
kers delimited the 10 and 90th percentiles, with
upper and lower dots representing 5 and 95th
percentile. Farming income was expressed as
percentage of total farmer household income.
Shown N fertilizer rates were per crop season.
Numbers at the top of the bottom panels in­
dicates the average plants per hill. Hill density
is not shown in SSU where rice is broadcasted.
See Fig. 1 for full province names.

similar biophysical background shown by Yuan et al. (2021), suggesting “historical” perspective based on their experience with improvements in
room to reduce pesticide load via better integrated pest management rice variety over past decades, rather than the influence of varietal se­
(Fig. 6). In relation to other factors, K fertilizer was identified as a source lection based on the varieties available to them for planting now.
of rice yield variation for many provinces, and with the same frequency However, Cassman (2001) found similar discrepancies between scien­
as N fertilizer, suggesting that K may be limiting current farmer yields tists’ perceptions and results of two independent on-farm studies con­
and contributing to low fertilizer-N use efficiency, in rice-based crop­ ducted in the 1990′ s in Southeast Asia, and pointed out the need for
ping systems in Southeast Asia as suggested in previous reports methodologies that can reliably establish key yield constraints in farmer
(Dobermann et al., 1998; Buresh et al., 2010; Singh et al., 2021). fields for AR&D prioritization. Hence, the farmer data-driven approach
There was a substantial discrepancy between experts’ opinion and can be used to re-orient AR&D investments and inform policy so that
results from this study about causes of Yg. For example, experts major yield constraints can be better addressed. Our approach is also
considered cultivar and crop damage as major causes of Yg, while our useful for ex-ante assessment of the potential impact of adoption of
analysis revealed nutrients and control of pests, and nutrients and yield-improving practices. For example, fields following the suite of
establishment as most important factors in rice and maize, respectively. management practices leading to highest productivity in each province
A possible reason for such a small influence of variety could be because had yields that were 77% and 63% of the Yp (or Yw) for rice and maize,
the national rice and maize research institutions run variety tests and respectively (compared to average Yg of 52% of the Yp (or Yw) for rice
limit the number of varieties available to farmers to those that perform and 43% of the Yp (or Yw) for maize). Overall, 17% of rice fields across
best. Thus, it is possible that experts responded to this question from an our study areas have closed their exploitable Yg, and only few for maize

9
G. Rizzo et al. Field Crops Research 297 (2023) 108942

Fig. 6. Conditional inference tree for rice (top) and maize (bottom) based on the pooled survey data in Indonesia (n = 1446 and 670 for rice and maize, respectively).
Yield is expressed as percentage of yield potential (irrigated crops) or water-limited yield potential (rainfed crops). In each boxplot, the box delimits the first and third
yield quartiles, while the horizontal line shows the median. Upper and lower whiskers represent the 10th and 90th percentiles, respectively. Also shown are the
means for the terminal node. Variables appearing in the conditional inference trees (their acronyms and units) were fertilizer N rate (N_fertilizer_rate; kg N ha− 1),
fertilizer K rate (K_fertilizer_rate; kg K ha− 1), number of urea splits (Urea_splits; times), timing of the second urea split (FTS2; number of weeks after transplanting),
number of insecticide applications (Insecticide_application; times).

(5%), highlighting the room that exists in most fields to increase yield opportunity for Indonesia considering the favorable socio-economic
via better agronomic management. If one scaled out the yield level of the context, with ca. 80% of farmers having good access to labor, inputs,
high-yield fields, weighted by the respective crop area in each province, and extension services. Given expected 17% and 63% increase in na­
to the entire rice and maize producing area in Indonesia, the additional tional rice and maize demand to 2040 (Robinson et al., 2015; Yuan et al.,
annual maize and rice output would amount to 17.2 and 5.7 MT, 2022) in relation to average national production between 2015 and
respectively, representing a 30% and 23% increase relative to current 2020, the production increase derived from our scenarios would allow
production (FAOSTAT, 2015–2020). Hence, progress towards closing Indonesia to become self-sufficient for rice by 2040, and even generate a
existing Yg using currently available technologies offers substantial rice surplus available for export, while greatly reducing pressure for

10
G. Rizzo et al. Field Crops Research 297 (2023) 108942

Fig. 7. Relative variable importance ranking for the influence of management factors on rice and maize yield gaps (% of Yp or Yw) for the pooled data across
Indonesia, as determined using random forest regression. Only the top 10 variables with highest variable importance are shown. FTS1 and FTS2: timing of first and
second urea split, respectively.

Fig. 8. (a) Causes of yield gaps in rice and maize in

Indonesia as identified via analysis of farmer data. (b) Rank
of yield constraints in rice as derived from consultation
with local researchers and extension agents. Yield con­
straints were grouped into five categories: nutrients
(amount, splits, and timing), pest control (fungicide and/or
insecticide applications, nematicide, weeding time and
type), establishment (time, plant density, and land prepa­
ration), cultivar (year of release and source), and damages
(yield losses due to insects, diseases, weeds, rats, etc.).

maize area expansion and imports. In contrast, failure to increase rice error and perceptions that are not necessarily aligned with reality, ul­
and maize production on existing cropland area will lead to further crop timately reducing efficiency, and impairing the capacity of developing
area expansion into low yield, high-risk environments, as it is the case of countries to meet their production goals on existing cropland while
relatively new rice systems in Sumatra and Kalimantan and upland conserving natural resources for future generations.
maize in Sulawesi. And much of this expansion would occur at expense
of converting natural ecosystems such as rainforests and wetlands. 5. Conclusions
Hence, closing the exploitable Yg in Indonesia is not only vital to meet
future domestic demand but also to prevent loss of natural ecosystems Our study describes an approach to identify causes of Yg for major
and the biodiversity it contains. food crops at national scale based on farmer-reported data using
Agronomic research is being challenged to become more efficient intensive rice and maize cropping systems in Indonesia as case-studies.
and impactful. Meeting this challenge would require innovative ap­ We found larger Yg for rice than maize and poor association between
proaches to accelerate the rate at which yield-improving factors are maize and rice yields. Socio-economic factors explained differences in
discovered and adopted by farmers. We believe that our data-driven Yg across crop-provinces, while agronomic factors, including nutrients,
approach based on producer data and a robust spatial framework to pest control, and establishment, explained field-to-field yield variation
guide data collection, can help increase the return to investment on within them. Results from our analyses are not consistent with percep­
AR&D programs. In our case, we note that a total of 1,147 surveys tions of local scientists and extension educators, who gave more
conducted at representative sites within major rice and maize producing emphasis to cultivars and damage due to biotic and abiotic factors than
areas in Indonesia allowed generation of information relevant to 9.7 and to on-farm yield constraints such as nutrient and agronomic manage­
4 M ha planted with rice and maize, respectively. And while results from ment factors. Our study provides a basis to prioritize investments on
our data-driven approach require field validation, and more thought on AR&D programs at national and sub-national level and inform policy.
how to combine them into viable technological packages (Andrade et al.,
2022a), they provide guidance to researchers, extensionists, and policy
makers about which factors to focus on when aiming to increase pro­ Declaration of Competing Interest
ductivity and profit of current crop systems. Without adoption of new
approaches to identify causes of Yg, as the one described in the present The authors declare that they have no known competing financial
study, prioritization of AR&D programs will rely too heavily on trial and interests or personal relationships that could have appeared to influence
the work reported in this paper.

11
G. Rizzo et al. Field Crops Research 297 (2023) 108942

Data availability Cassman, K.G. , A. Dobermann , D.T. Walters , and H. Yang, 2003, Meeting cereal
demand while protecting natural resources and improving environmental quality.
Annual Review of Environment and Resources 28 (1): 315–358.
Data will be made available on request. Clapp, J., 2017. Food self-sufficiency: Making sense of it, and when it makes sense. Food
Policy 66, 88–96.
Acknowledgments Cock, J., Oberthür, T., Isaacs, C., Läderach, P.R., Palma, A., Carbonell, J., Victoria, J.,
Watts, G., et al., 2011. Crop management based on field observations: Case studies in
sugarcane and coffee. Agric. Syst. 104 (9), 755–69.
We thank the Indonesian farmers who participated in this study and Dawe, D., Timmer, C.P., 2012. Why stable food prices are a good thing: Lessons from
their willingness to provide data from their fields. Support for this stabilizing rice prices in Asia. Glob. Food Secur. 1 (2), 127–33.
Desiere, S., Jolliffe, D., 2018. Land productivity and plot size: Is measurement error
project comes from the Robert B. Daugherty Water for Food Global driving the inverse relationship? J. Dev. Econ. 130, 84–98.
Institute at the University of Nebraska-Lincoln (UNL), the Indonesian Dobermann, A., Fairhurst, T., 2000. Rice: Nutrient Disorders & Nutrient Management. 1. ed.
Agency for Agricultural Instrument Standardization (BSIP), and the Potash & Phosphate Institute,, Signapore.
Dobermann, A., Cassman, K.G., Mamaril, C.P., Sheehy, J.E., 1998. Management of
Office of Global Engagement at the Institute of Agriculture and Natural phosphorus, potassium, and sulfur in intensive, irrigated lowland rice. Field Crops
Resources at UNL. The authors thank Syafrudin, Nurwulan Agustiani, Res. 56 (1–2), 113–38.
Evert Y. Hosang, Fitrah Irawan, Andi Yulyani Fadwiwati from the Na­ Dobermann, A., Witt, C., Dawe, D. (Eds.), 2004. Increasing productivity of intensive rice
systems through site-specific nutrient management. International Rice Research
tional Research and Innovation Agency (BRIN), Leonardus Krisnadi, Institute, Philippines.
Andarias Makka, Aulia Dina Pramesti, and Hiryana Windiyani from BSIP Dobermann, A., Witt, C., Dawe, D., Abdulrachman, S., Gines, H.C., Nagarajan, R.,
for their assistance in data collection. We also thank Dr Shen Yuan and Satawathananont, S., Son, T.T., et al., 2002. Site-specific nutrient management for
intensive rice cropping systems in Asia. Field Crops Res. 74 (1), 37–66.
Dr Nanyan Deng (Huazhong Agricultural University, China) for their
Evans, L.T., 1993. Crop Evolution, Adaptation and Yield. Cambridge University Press,,
help on estimating future demand for rice and maize in Indonesia and Cambridge, UK.
previous rice simulations and Mrs. Ngu Kah Hui (UNL) for assistance to FAOSTAT Production Data (FAO, accessed 18 May 2022); www.fao.org/faostat/en/
input survey data into digital format. #data.
Farmaha, B.S., Lobell, D.B., Boone, K.E., Cassman, K.G., Yang, H.S., Grassini, P., 2016.
Contribution of persistent factors to yield gaps in high-yield irrigated maize. Field
Appendix A. Supporting information Crops Res. 186,, 124–132.
Fraval, S., Hammond, J., Wichern, J., Oosting, S.J., Boer, I.J.M.D., Teufel, N.,
Lannerstad, M., Waha, K., et al., 2019. Making the most of imperfect data: a critical
Supplementary data associated with this article can be found in the evaluation of standard information collected in farm household surveys. Exp. Agric.
online version at doi:10.1016/j.fcr.2023.108942. 55 (2), 230–250.
Gourlay, S., Kilic, T., Lobell, D., 2017. Could the Debate Be Over? Errors in Farmer-
Reported Production and Their Implications for the Inverse Scale-Productivity
References Relationship in Uganda. World Bank,, Washington, DC.
Grassini, P., Thorburn, J., Burr, C., Cassman, K.G., 2011. High-yield irrigated maize in
Agus, F., Andrade, J.F., Rattalino Edreira, J.I., Deng, N., Purwantomo, D.K.G., the Western U.S. Corn Belt: I. On-farm yield, yield potential, and impact of
Agustiani, N., Aristya, V.E., Batubara, S.F., et al., 2019. Yield gaps in intensive rice- agronomic practices. Field Crops Res. 120 (1), 142–150.
maize cropping sequences in the humid tropics of Indonesia. Field Crops Res. 237, Grassini, P., van Bussel, L.G.J., Van Wart, J., Wolf, J., Claessens, L., Yang, H.,
12–22. Boogaard, H., de Groot, H., et al., 2015a. How good is good enough? Data
Agustiani, N., Deng, N., Edreira, J.I.R., Girsang, S.S., Syafruddin, Sitaresmi, T., requirements for reliable crop yield simulations and yield-gap analysis. Field Crops
Pasuquin, J.M.C., Agus, F., et al., 2018. Simulating rice and maize yield potential in Res. 177, 49–63.
the humid tropical environment of Indonesia. Eur. J. Agron. 101, 10–19. Grassini, P., Torrion, J.A., Yang, H.S., Rees, J., Andersen, D., Cassman, K.G., Specht, J.E.,
Andrade, J.F., Mourtzinis, S., Rattalino Edreira, J.I., Conley, S.P., Gaska, J., Kandel, H.J., 2015b. Soybean yield gaps and water productivity in the western U.S. Corn Belt.
Lindsey, L.E., Naeve, S., et al., 2022a. Field validation of a farmer supplied data Field Crops Res. 179, 150–63.
approach to close soybean yield gaps in the US North Central region. Agric. Syst. Hothorn, T., Zeileis, A., 2015. partykit: A Modular Toolkit for Recursive Partytioning in
200, 103434. R. J. Mach. Learn. Res. 16 (118), 3905–3909.
Andrade, J.F., Cassman, K.G., Rattalino Edreira, J.I., Agus, F., Bala, A., Deng, N., Hunter, M.C., Smith, R.G., Schipanski, M.E., Atwood, L.W., Mortensen, D.A., 2017.
Grassini, P., 2022b. Impact of urbanization trends on production of key staple crops. Agriculture in 2050: Recalibrating Targets for Sustainable Intensification. BioScience
Ambio 51 (5), 1158–67. 67 (4), 386–91.
Arifin, B., N. Achsani, A. Drajat, M. Karlina, S. Ahmad, H. Firdaus, and Jakarta, 2018, IFPRI, 2017, IMPACT Projections of Food Production, Consumption, and Hunger to
Modeling the Future of Indonesian Food Consumption: Final Report. 2050, With and Without Climate Change: Extended Country-level Results for 2017
Bates, D., Mächler, M., Bolker, B., Walker, S., 2015. Fitting linear mixed-effects models GFPR Annex Table 6. Translated by Institutions CGIAR Research Program on
using lme4. J. Stat. Softw. 67, 1–48. Policies, Agriculture and Food Security (CCAFS) CGIAR Research Program on
Boling, A., Tuong, T.P., Jatmiko, S.Y., Burac, M.A., 2004. Yield constraints of rainfed Climate Change, and Bill and Melinda Gates Foundation (BMGF). Harvard
lowland rice in Central Java, Indonesia. Field Crops Res. 90 (2–3), 351–60. Dataverse.
Boling, A.A., Tuong, T.P., van Keulen, H., Bouman, B.A.M., Suganda, H., Spiertz, J.H.J., International Food Policy Research Institute, 2019, Global Spatially-Disaggregated Crop
2010. Yield gap of rainfed rice in farmers’ fields in Central Java, Indonesia. Agric. Production Statistics Data for 2010 Version 2.0. Edited by International Food Policy
Syst. 103 (5), 307–15. Research Institute. Translated by CGIAR Platform for Big Data in Agriculture,
Brankov, T., Matkovski, B., Jeremić, M., Đurić, I., 2021. Food Self-Sufficiency of the SEE Institutions CGIAR Research Program on Policies and Markets, The Bill and Melinda
Countries; Is the Region Prepared for a Future Crisis? Sustainability. Gates Foundation, NASA Harvest, USAID Feed the Future Innovation Lab for
Multidisciplinary Digital Publishing Institute,, p. 8747. Collaborative Research on Sustainable Intensification, and USAID Feed the Future
Buresh, R.J., 2010. Nutrient best management practices for rice, maize, and wheat in Innovation Lab for Small Scale Irrigation. Harvard Dataverse.
Asia. World Congr. Soil Sci. Vol. 16. International Rice Research Institute, ed. 1984, Terminology for rice growing
Buresh, R.J., Pampolino, M.F., Witt, C., 2010. Field-specific potassium and phosphorus environments. Los Baños, Laguna, Philippines: International Rice Research Institute.
balances and fertilizer requirements for irrigated rice-based cropping systems. Plant van Ittersum, M.K., Rabbinge, R., 1997. Concepts in production ecology for analysis and
Soil 335 (1–2), 35–64. quantification of agricultural input-output combinations. Field Crops Res. 52,
Buresh, R.J., Correa, T.Q., Pabuayon, I.L.B., Laureles, E.V., Choi, I.-R., 2021. Yield of 197–208.
irrigated rice affected by asymptomatic disease in a long-term intensive Kendall, M.G., 1938. A new measure of rank correlation. Biometrika 30 (1–2), 81–93.
monocropping experiment. Field Crops Res. 265, 108121. Kriewald, S., Pradhan, P., Costa, L., Ros, A.G.C., and Kropp, J.P., 2019, Hungry cities:
van Bussel, L.G.J., Grassini, P., Van Wart, J., Wolf, J., Claessens, L., Yang, H., how local food self-sufficiency relates to climate change, diets, and urbanisation.
Boogaard, H., de Groot, H., et al., 2015. From field to atlas: Upscaling of location- Environmental Research Letters 14 (9). IOP Publishing: 094007.
specific yield gap estimates. Field Crops Res. 177, 98–108. Laborte, A.G., de Bie, C.A.J.M., Smaling, E.M.A., Moya, P.F., Boling, A.A., Van
Cassman, K.G., 1999. Ecological intensification of cereal production systems: Yield Ittersum, M.K., 2012. Rice yields and yield gaps in Southeast Asia: Past trends and
potential, soil quality, and precision agriculture. Proc. Natl. Acad. Sci. USA 96, future outlook. Eur. J. Agron. 36 (1), 9–20.
5952–5959. Liaw, A., Wiener, M., 2002. Classification and Regression by randomForest. R. N. 2 (3),
Cassman, K.G., 2001, Crop science research to assure food security. In Crop science: 18–22.
progress and prospects. Papers presented at the Third International Crop Science Lobell, D.B., Ortiz-Monasterio, J.I., Lee, A.S., 2010. Satellite evidence for yield growth
Congress, Hamburg, Germany, 17–22 August 2000, ed. J. Nösberger, H. H. Geiger, opportunities in Northwest India. Field Crops Res. 118 (1), 13–20.
and P. C. Struik, 33–51. Wallingford: CABI. Lobell, D.B., Azzari, G., Burke, M., Gourlay, S., Jin, Z., Kilic, T., Murray, S., 2020. Eyes in
Cassman, K.G., Grassini, P., 2020. A global perspective on sustainable intensification the Sky, Boots on the Ground: Assessing Satellite- and Ground-Based Approaches to
research. Nat. Sustain. 3 (4), 262–268. Crop Yield Measurement and Analysis. Am. J. Agric. Econ. 102 (1), 202–19.

12
G. Rizzo et al. Field Crops Research 297 (2023) 108942

Lollato, R.P., Figueiredo, B.M., Dhillon, J.S., Arnall, D.B., Raun, W.R., 2019. Wheat grain Singh, V.K., Dwivedi, B.S., Rathore, S.S., Mishra, R.P., Satyanarayana, T., Majumdar, K.,
yield and grain-nitrogen relationships as affected by N, P, and K fertilization: A 2021. Timing Potassium Applications to Synchronize with Plant Demand, 363–84.
synthesis of long-term experiments. Field Crops Res. 236, 42–57. In: Murrell, T.S., Mikkelsen, R.L., Sulewski, G., Norton, R., Thompson, M.L. (Eds.),
Mourtzinis, S., Rattalino Edreira, J.I., Grassini, P., Roth, A.C., Casteel, S.N., Ciampitti, I. Improving Potassium Recommendations for Agricultural Crops. Springer International
A., Kandel, H.J., Kyveryga, P.M., et al., 2018. Sifting and winnowing: Analysis of Publishing, Cham.
farmer field data for soybean in the US North-Central region. Field Crops Res. 221, So, H.B., Ringrose-Voase, A.J., 2000. Management of clay soils for rainfed lowland rice-
130–41. based cropping systems: an overview. Soil Tillage Res. 56 (1), 3–14.
Munaro, L.B., Hefley, T.J., DeWolf, E., Haley, S., Fritz, A.K., Zhang, G., Haag, L.A., Stuart, A.M., Pame, A.R.P., Silva, J.V., Dikitanan, R.C., Rutsaert, P., Malabayabas, A.J.B.,
Schlegel, A.J., et al., 2020. Exploring long-term variety performance trials to Lampayan, R.M., Radanielson, A.M., et al., 2016. Yield gaps in rice-based farming
improve environment-specific genotype × management recommendations: A case- systems: Insights from local studies and prospects for future analysis. Field Crops
study for winter wheat. Field Crops Res. 255, 107848. Res. 194, 43–56.
Pasuquin, J.M., Pampolino, M.F., Witt, C., Dobermann, A., Oberthür, T., Fisher, M.J., Stuart, A.M., Devkota, K.P., Sato, T., Pame, A.R.P., Balingbing, C., My Phung, N.T.,
Inubushi, K., 2014. Closing yield gaps in maize production in Southeast Asia through Kieu, N.T., Hieu, P.T.M., et al., 2018. On-farm assessment of different rice crop
site-specific nutrient management. Field Crops Res. 156, 219–30. management practices in the Mekong Delta, Vietnam, using sustainability
Rattalino Edreira, J.I., Mourtzinis, S., Conley, S.P., Roth, A.C., Ciampitti, I.A., Licht, M. performance indicators. Field Crops Res. 229, 103–14.
A., Kandel, H., Kyveryga, P.M., et al., 2017. Assessing causes of yield gaps in Swastika, D.K., F. Kasim, K. Suhariyanto, W. Sudana, R. Hendayana, R.V. Gerpacio, and
agricultural areas with diversity in climate and soils. Agric. For. Meteorol. 247, P.L. Pingali. 2004. Maize in Indonesia: production systems, constraints and research
170–80. priorities. CIMMYT.
Rattalino Edreira, J.I., Mourtzinis, S., Azzari, G., Andrade, J.F., Conley, S.P., Specht, J.E., Tenorio, F.A.M., McLellan, E.L., Eagle, A.J., Cassman, K.G., Torrion, J.A., Grassini, P.,
Grassini, P., 2020. Combining field-level data and remote sensing to understand 2021. Disentangling management factors influencing nitrogen balance in producer
impact of management practices on producer yields. Field Crops Res. 257, 107932. fields in the western Corn Belt. Agric. Syst. 193, 103245.
Robinson, S., Mason-D′ Croz, D., Sulser, T., Islam, S., Robertson, R., Zhu, T., Gueneau, A., United Nations, D. of E. and S.A., 2019, World Population Prospects 2019, Online
Pitois, G., et al., 2015. The International Model for Policy Analysis of Agricultural Edition. Rev. 1.
Commodities and Trade (IMPACT): Model Description for Version 3. SSRN Electron. Wade, L.J., Amarante, S.T., Olea, A., Harnpichitvitaya, D., Naklang, K., Wihardjaka, A.,
J. Sengar, S.S., Mazid, M.A., et al., 1999. Nutrient requirements in rainfed lowland rice.
Rossini, M.A., Maddonni, G.A., Otegui, M.E., 2011. Inter-plant competition for resources Field Crops Res. 64 (1), 91–107.
in maize crops grown under contrasting nitrogen supply and density: Variability in van Wart, J., van Bussel, L.G.J., Wolf, J., Licker, R., Grassini, P., Nelson, A., Boogaard, H.,
plant and ear growth. Field Crops Res. 121 (3), 373–80. Gerber, J., et al., 2013. Use of agro-climatic zones to upscale simulated crop yield
Silva, J.V., Reidsma, P., Lourdes Velasco, Ma, Laborte, A.G., van Ittersum, M.K., 2018. potential. Field Crops Res. 143, 44–55.
Intensification of rice-based farming systems in Central Luzon, Philippines: Yuan, S., B.A. Linquist, L.T. Wilson, K.G. Cassman, A.M. Stuart, V. Pede, B. Miro, K. Saito,
Constraints at field, farm and regional levels. Agric. Syst. 165, 55–70. et al., 2021, Sustainable intensification for a larger global rice bowl. Nature
Silva, J.V., Pede, V.O., Radanielson, A.M., Kodama, W., Duarte, A., de Guia, A.H., Communications 12 (1). Nature Publishing Group: 7163.
Malabayabas, A.J.B., Pustika, A.B., et al., 2022. Revisiting yield gaps and the scope Yuan, S., A.M. Stuart, A.G. Laborte, J.I. Rattalino Edreira, A. Dobermann, L.V.N. Kien, L.
for sustainable intensification for irrigated lowland rice in Southeast Asia. Agric. T. Thúy, K. Paothong, et al., 2022, Southeast Asia must narrow down the yield gap to
Syst. 198, 103383. continue to be a major rice bowl. Nature Food 3 (3). Nature Publishing Group:
217–26.

13