Towards an ecient and risk aware strategy for guiding farmers in identifying best crop management

2025-04-15 0 0 8.4MB 31 页 10玖币

侵权投诉

Towards an eﬃcient and risk aware strategy for guiding

farmers in identifying best crop management

Romain Gautron ∗1,2,3, Dorian Baudry4, Myriam Adam5,6,7, Gatien N. Falconnier1,2,8, and

Marc Corbeels1,2,9

1AIDA, Univ Montpellier, France.

2CIRAD, Montpellier, France.

3CGIAR Platform for Big Data in Agriculture, Alliance of Bioversity International and CIAT, Km 17, Recta Cali

Palmira 763537, Colombia.

4Univ. Lille, CNRS, Inria, Centrale Lille, UMR 9198-CRIStAL, F-59000 Lille, France.

5CIRAD, UMR AGAP Institut, Bobo-Dioulasso 01, Burkina Faso.

6UMR AGAP Institut, Univ Montpellier, CIRAD, INRAE, Institut Agro, Montpellier, France.

7Institut National de l’Environnement et de Recherches Agricoles (INERA), Burkina Faso.

8International Maize and Wheat Improvement Centre (CIMMYT)-Zimbabwe, 12.5 km Peg Mazowe Road, Harare,

Zimbabwe.

9International Institute of Tropical Agriculture, PO Box 30772, Nairobi, 00100, Kenya.

October 11, 2022

Abstract

Identiﬁcation of best performing fertilizer practices among a set of contrasting practices with ﬁeld trials

is challenging as crop losses are costly for farmers. To identify best management practices, an “intuitive

strategy” would be to set multi-year ﬁeld trials with equal proportion of each practice to test. Our objective

was to provide an identiﬁcation strategy using a bandit algorithm that was better at minimizing farmers’

losses occurring during the identiﬁcation, compared with the “intuitive strategy”. We used a modiﬁcation

of the Decision Support Systems for Agro-Technological Transfer (DSSAT) crop model to mimic ﬁeld trial

responses, with a case-study in Southern Mali. We compared fertilizer practices using a risk-aware measure,

the Conditional Value-at-Risk (CVaR), and a novel agronomic metric, the Yield Excess (YE). YE accounts

for both grain yield and agronomic nitrogen use eﬃciency. The bandit-algorithm performed better than the

intuitive strategy: it increased, in most cases, farmers’ protection against worst outcomes. This study is a

methodological step which opens up new horizons for risk-aware ensemble identiﬁcation of the performance

of contrasting crop management practices in real conditions.

1 Introduction

Identifying site-speciﬁc best-performing crop management is crucial for farmers to increase their income

from crop production, but also for minimizing the negative environmental impact of cropping activities (Tilman

et al.,2002). However, due to weather variability, the identiﬁcation of these practices can be challenging, in

particular with rainfed farming: what worked best in a wet year, might not work in the next season, when

rainfall is less (Aﬀholder,1995). In fact, the performance of crop management at a given site has an underlying

“hidden” distribution due to inter-annual weather variability, thus creating great uncertainty (Fosu-Mensah

et al.,2012). Because crop management decisions are recurrent, i.e. they are repeated for each new crop

growing season, the identiﬁcation of optimal crop management falls into the category of sequential decision

making under uncertainty (Gautron et al.,2022). Computer-based decision support tools can allow farmers to

make more informed (less uncertain) decisions about their cropping practices from one year to the next, and can

facilitate farmers’ risk management in the face of seasonal weather variability (Hochman and Carberry,2011).

∗romain.gautron@cirad.fr

arXiv:2210.04537v1 [cs.AI] 10 Oct 2022

There exist numerous decision support tools of widely ranging complexity for crop management, introduced to

farmers with varying degrees of success (Gautron et al.,2022).

Machine learning (ML) and more generally artiﬁcial intelligence (AI) can help address sequential decision

making under uncertainty. In particular, the bandit algorithm paradigm (Lattimore and Szepesv´ari,2020)

considers a decision-maker, called agent, who repeatedly faces a choice between contending actions, and has to

iteratively improve its decisions with trials. The canonical bandit problem originates from clinical trials with

sequential drug allocation (Thompson,1933). At each time step, the agent chooses one action (i.e., one drug

for a patient) amongst a set of possible actions. Each action provides a reward (i.e.; tumor cell reduction after

taking the drug), drawn from a corresponding unknown reward distribution (i.e., the distribution of tumor cell

reduction for the drug). The optimal action has the reward distribution with the highest mean reward (i.e.,

the highest mean tumor cell reduction). The objective of the agent is to sequentially choose actions such that

the expected sum of rewards is maximized. Maximizing the total expected rewards is equivalent to minimizing

the regret, which is a measure of the total losses that occur with sub-optimal actions (Robbins,1952).

Iteratively, the agent reﬁnes his next decision based on all previous results. To know how a given action

performs, a suﬃcient number of (possibly poor) rewards is required: this is the exploration phase. To

maximize the expected sum of rewards, the previous actions that provided good results so far must be selected

more frequently; this is the exploitation phase. Bandit algorithms aim at ﬁnding the right balance between

exploration and exploitation. This exploration-exploitation dilemma is a reality for farmers when implementing

crop management. Farmers typically want to minimize overall crop yield losses and typically explore the

performance of promising new crop management practices on small test plots (Cerf and Meynard,2006;Evans

et al.,2017). They avoid potentially large crop yield losses from new management by managing a gradual

transition between the current management and the promising new one(s), based on the results they obtain on

the small test plots.

The objective of this paper is to develop a novel strategy to identify best crop management. We set as

baseline an “intuitive strategy” which consists in identifying the best crop management through multi-year

ﬁeld trials in which a set of crop management practices is tested in an equiproportional way. We compare this

“intuitive strategy” to a novel crop management identiﬁcation strategy, based on a bandit algorithm. This

novel identiﬁcation strategy aims to minimize farmers’ yield losses occurring during the identiﬁcation process,

compared to the intuitive strategy. Thus, we test the hypothesis that bandit algorithm can help farmers to

better identify the best crop management for their context, while further minimizing crop yield losses related

to sub-optimal choices in new crop management.

Our case study considers the rainfed maize production in southern Mali, and we compare the performance

of both crop management identiﬁcation strategies based on maize growth simulations using a calibrated crop

model in order to mimic real-world performance of crop management. The novel identiﬁcation strategy does,

however, not depend on model simulations, and ultimately aims at being applied in real ﬁeld conditions. As

for crop management, we focus on nitrogen fertilization. Tailoring nitrogen fertilizer recommendations to

farmers’ contexts is known to be challenging. Indigenous soil nitrogen supply, depending to a large extent

on past-season events, is not accurately known to farmers, whilst in-season nitrogen mineralization depends

largely on weather events(Morris et al.,2018), themselves uncertain. Crop nitrogen requirements, such as with

maize, are related to speciﬁc crop growth stages (Hanway,1963) and excessive mineral nitrogen supply can

induce nitrate leaching, especially in wet conditions (Meisinger and Delgado,2002). Therefore, there are a

priori no upfront optimal nitrogen fertilizer practices.

2 Methods

2.1 Virtual crop management identiﬁcation problem

In our virtual crop management identiﬁcation problem, a population or ensemble of farmers joined a

participatory experiment to identify the best nitrogen fertilizer practices for maize production in their region,

Koutiala in southern Mali. A total population of 500 farmers was considered. The distribution of soil types of

the ﬁelds associated with the group of farmers was representative of the region (Table 1). A total population of

500 farmers was considered. Each farmer belonged to a cohort that corresponded to an ensemble of farmers

growing maize on the same soil type. For each cohort, we wanted to identify the best nitrogen fertilizer practice

from a set of recommended practices (see Table 3and Section 2.1.1 for the performance metrics we considered).

cohort 1

cohort 2

identiﬁcation

strategy for

cohort 1

identiﬁcation

strategy for

cohort 2

season

volunteers

...

farmer

population

beginning of the season identiﬁcation process

Figure 1: Yearly process to generate nitrogen fertilizer recommendations: at the beginning of the crop ping

season. Individuals from the overall farmer population volunteered to test a fertilizer practice. Similar symbols

represent a cohort, i.e., a group of farmers having ﬁelds with the same soil type. The group of volunteer farmers

was broken down by cohort and researchers independently generated fertilizer recommendations for each cohort.

Researchers did not control the number of volunteers from the respective cohorts In this example, only three of

the four possible cohorts are found in the volunteer group.

The research team set the additional objective to limit the crop yield losses of individual farmers that could

arise from poor nitrogen fertilizer practice recommendations during the identiﬁcation process.

At the beginning of each crop growing season, we assumed that a random number of farmers (uniformly

obtained between 250 and 350) of the population of 500 farmers volunteered to apply the recommended fertilizer

applications provided by the research team. Each year, the group of volunteers was variable in size and in the

representation of cohorts, as could occur in reality (Figure 1). Thus, researchers did not control the composition

of the group of volunteers. Each farmer indicated the ﬁelds and corresponding soils on which she/he planned

to grow maize. Researchers then provided a fertilizer recommendation (Table 3) to each farmer for the ongoing

season, depending on her/his soil i.e. cohort. At the end of the season, volunteer farmers shared their results in

terms of maize grain yields with the research team, allowing to reﬁne the recommendations for the next season.

The whole process was repeated during 20 consecutive years following the same process (Figure 2a).

Nitrogen fertilizer practices.

Ten nitrogen fertilizer practices were considered as recommendations in the

virtual modeling experiment (see Table 2). Practices 0 to 7 explored the following set of split applications for a

total amount of 135 kg N/ha applied:

- Two split applications (practice 0): 15 days after planting (DAP) and 30 DAP.

- Three split applications (practice 4) :15 DAP, 30 DAP and 45 DAP.

Split applications according to the rainfall amount (practices 2, 3 and 6, 7): 2nd and 3rd top-dressing

applications only if the cumulated rainfall amount from the start of the season to 30 DAP exceeds the

30th percentile of historical rainfall i.e. 200 mm.

Split applications according to plant nitrogen content (practices 1, 3 and 5, 7): 2nd and 3rd top-dressing

applications only if the simulated nitrogen stress factor (

NSTRES

in DSSAT, see below) exceeds 0.2 (0 no

stress, 1 maximal stress), hereby mimicking the use of a portable chlorophyll meter to monitor plant

nitrogen content (e.g. Kalaji et al.,2017).

Practice 8 corresponded to the optimal fertilization for maize (70 kg N/ha) in the study area based on

simulations (Huet et al.,2022) , i.e. the average of the N fertilizer rates that were observed to result in

maximum positive return on fertilizer investment (Getnet et al.,2016). Finally, practice 9 (180 kg N/ha)

corresponded to a nitrogen fertilizer practice that is likely excessive. For all these practices, the nitrogen

fertilizer applied was assumed to be ammonium nitrate broadcasted on the soil surface.

For Tyears:

get volunteer

farmers for

current year

farmer

population

get all

volunteers’

results

experts’

identiﬁcation

strategy

assign

fertilizer

practices to

all volunteers

beginning of the season

end of the

season

year ←year + 1

(a) Diagram of the ensemble best fertilizer identiﬁcation

process. Each year, a group of volunteer farmers test

fertilizer practices recommended by experts and contribute

to identifying the best fertilizer practices for the region.

At the end of each season, the farmers share their results

with experts. The experts will use these results to improve

their recommendations for the next growing season. The

process repeats for a total number of Tyears.

For Ttimes:

choose an

action

from

actions

observe an

uncertain result

of action

make the

action

t←t+ 1

(b) Canonical bandit problem. For

times, an agent

sequentially makes a decision on an action

from the set

{

,· · · , K}

of possible actions. After making the action

, the agent observes an uncertain result

. This result is

sampled from a ﬁxed distribution, unknown to the agent,

which corresponds to the eﬀect of action kt.

Figure 2: Schematic representation of the ensemble best fertilization identiﬁcation process and the canonical

bandit problem.

Table 1: Main properties of the soil types of the ﬁelds of farmers growing maize in Koutiala, Mali (Adam

et al.,2020). ‘

SLOC

.’ stands for soil organic matter (g C/ 100 g soil, mean value for the 0-30 cm topsoil); ‘

SLDR

’

stands for soil drainage rate (fraction/day); ‘

SLDP

’ stands for soil depth (cm); ‘Prop’ stands for the percentage

of each soil type present in the study area.

Soil name Texture SLDR SLOC SLDP Prop.

ITML840101 clay loam 0.60 0.20 110 7%

ITML840102 loam 0.60 0.45 100 9%

ITML840103 silty loam 0.60 0.27 160 21%

ITML840104

silty clay loam

0.25 0.70 105 4%

ITML840105

silty clay loam

0.40 0.35 120 24%

ITML840106 loam 0.60 0.30 110 27%

ITML840107

silty clay loam

0.25 0.60 105 8%

Table 2: Maize nitrogen fertilizer recommendations for maize in Koutiala, Southern Mali, that were considered

in the virtual experiment. Whether or not rainfall and plant nitrogen stress were considered as factors for the

fertilizer recommendation is indicated by Yes or No. ‘

NSTRES

’ stands for plant nitrogen stress and ‘DAP’ for

days after planting.

index max

amount

applied

(kgN/ha)

max

applica-

tions

rainfall

thresh-

old

NSTRES

thresh-

old

15 DAP N

(kgN/ha)

30 DAP N

(kgN/ha)

45 DAP N

(kgN/ha)

0 135 2 No No 15 120 0

1 135 2 No Yes 15 120 0

2 135 2 Yes No 15 120 0

3 135 2 Yes Yes 15 120 0

4 135 3 No No 15 60 60

5 135 3 No Yes 15 60 60

6 135 3 Yes No 15 60 60

7 135 3 Yes Yes 15 60 60

8 70 2 No No 23 0 47

9 180 3 No No 60 60 60

Maize growth simulations.

In order to get a proxy for real-world performances of the maize nitrogen

fertilizer practices, we simulated maize growth responses to fertilization under the growing conditions of Koutiala

in southern Mali using

gym-DSSAT

(Gautron and Padr´on Gonz´alez,2022).

gym-DSSAT

is a modiﬁcation of

the DSSAT crop simulator (Hoogenboom et al.,2019) to allow a user to read DSSAT internal states and

take daily fertilization decisions during the simulations (e.g. based on DSSAT internal states). For each soil

type in Table 1that was parametrized in DSSAT using the data from Adam et al. (2020), each simulated

maize grain yield value is a sample of the response distribution for the considered fertilizer practice. This

response distribution is the result of weather variability, generated in our study by the stochastic weather

generator WGEN (Richardson and Wright,1984;Soltani and Hoogenboom,2003), which was calibrated using

the 47-year-long weather records from N’tarla, about 30 km from Koutiala (Ripoche et al.,2015). The ‘sotubaka’

maize cultivar (from the DSSAT default cultivar list) was used for all model simulations as a representative

of maize variety in southern Mali. Water and nitrogen stresses were simulated, but yield reduction through

pests and diseases were not considered, neither was weed competition. In the model simulations, a diﬀerent

weather time series was generated for each growing season and for each recommendation using WGEN, inducing

independent simulated maize yield responses to nitrogen fertilization. Section Aof Supplementary Materials

gives further details of the simulation settings.

We simulated 10

times the maize grain yield responses to a given fertilizer practice for the diﬀeret soil

types, which corresponds to 10

hypothetical growing seasons. These samples were used i) to ensure that

simulated maize yield responses were in realistic expected ranges, ii) to qualitatively evaluate the complexity of

the decision problem, and iii) to determine best nitrogen fertilizer practices whilst analyzing the performance

of the crop management identiﬁcation strategies. The samples were not provided to the algorithms prior to

their application (i.e. no prior knowledge of the problem).

2.1.1 Performance indicators of fertilizer practices

A criterion to evaluate both the economic and environmental performance of a fertilizer practice

Agronomic Nitrogen use Eﬃciency (ANE), as deﬁned in Vanlauwe et al. (2011):

ANEπ:= Yπ−Y0

Nπ(1)

where

Yπ

is the crop yield obtained with the nitrogen fertilizer practice

which required a quantity

Nπ

nitrogen and

is the yield of the control obtained in the same conditions without nitrogen fertilization.

Maximising ANE is a proxy of minimizing the quantity of nitrogen losses, e.g. through nitrate leaching.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

TowardsanecientandriskawarestrategyforguidingfarmersinidentifyingbestcropmanagementRomainGautron*1,2,3,DorianBaudry4,MyriamAdam5,6,7,GatienN.Falconnier1,2,8,andMarcCorbeels1,2,91AIDA,UnivMontpellier,France.2CIRAD,Montpellier,France.3CGIARPlatformforBigDatainAgriculture,AllianceofBioversityInternati...

展开>> 收起<<

Towards an ecient and risk aware strategy for guiding farmers in identifying best crop management.pdf

共31页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Towards an ecient and risk aware strategy for guiding farmers in identifying best crop management

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: