J. Dö e al.: Digi ale In as uk u en,
Lec u e No es in In o ma ics (LNI), Gesellscha ü In o ma ik, Bonn 2025 321
Da a-d i en ni ogen managemen
Le e aging his o ical da a and machine lea ning o economic op ima
Cus odio Ma a el
1
and And eas Meye -Au ich
2
Abs ac : E icien ni ogen managemen is c ucial o economic and en i onmen al sus ainabili y
in ag icul u e. Excessi e ni ogen use leads o pollu ion and highe cos s, while insu icien applica-
ion educes c op yields and p o i abili y. This s udy builds on exis ing ni ogen esponse models by
using machine lea ning o op imize ni ogen use, le e aging da a om long- e m expe imen s. The
da ase includes ni ogen ea men le els, c op yields, and wea he da a, enabling a comp ehensi e
analysis o ni ogen managemen p ac ices. By employing Suppo Vec o Reg ession (SVR) and
Random Fo es (RF) models, his esea ch cap u es non-linea ela ionships be ween wea he a ia-
bles and Economic Op imum Ni ogen Ra e (EONR). Resul s show RF models o ou pe o m SVR,
wi h lowe Roo Mean Squa ed E o (RMSE) and highe R2, highligh ing RF’s obus ness in man-
aging a iabili y. The s udy also e alua es he e ec s o o ganic e iliza ion, emphasizing he po-
en ial o da a-d i en ni ogen managemen o suppo sus ainable ag icul u e wi h localized, p ecise
ecommenda ions.
Keywo ds: ni ogen managemen , machine lea ning, EONR, long- e m ield expe imen s, NUE
1 In oduc ion
E icien ni ogen managemen is essen ial o sus ainable ag icul u e, balancing economic
and en i onmen al ou comes [Zh15]. Excessi e ni ogen use inc eases cos s and con ib-
u es o wa e pollu ion and g eenhouse gas emissions, while insu icien applica ion e-
duces c op yields, impac ing p o i abili y and ood p oduc ion [Ti02]. Op imizing ni ogen
use, he e o e, balances economic bene i s wi h en i onmen al s ewa dship.
T adi ional ni ogen managemen models, o en gene alized o maximize yields, may no
adequa ely cap u e he complex in e ac ions be ween ni ogen applica ion, soil condi ions,
and en i onmen al ac o s equi ed o op imal p o i abili y [Mo18; LS03]. The Economic
1
Leibniz Ins i u e o Ag icul u al Enginee ing and Bioeconomy (ATB), Max-Ey h-Allee 100, 14469, Po s-
dam, Ge many, cma a el@a b-po sdam.de, h ps://o cid.o g/0000-0002-3800-7887
2
Leibniz Ins i u e o Ag icul u al Enginee ing and Bioeconomy (ATB), Max-Ey h-Allee 100, 14469, Po s-
dam, Ge many, AMeye -Au ich@a b-po sdam.de, h ps://o cid.o g/0000-0002-8235-0703
322 Cus odio Ma a el and And eas Meye -Au ich
Op imum Ni ogen Ra e (EONR) o e s a mo e e ined app oach by de e mining he ni-
ogen a e ha maximizes economic e u ns, p o iding a me s wi h ailo ed ecommen-
da ions.
This s udy uses machine lea ning o ad ance ni ogen managemen p ac ices, cap u ing
non-linea ela ionships be ween ni ogen use e iciency (NUE), wea he pa e ns, and
ea men le els [La23; We22]. Unlike con en ional models, his app oach emphasizes he
in e ac ion be ween ni ogen a es and wea he a iables, p o iding localized, p ecise ec-
ommenda ions o speci ic ield condi ions. In eg a ing his o ical and eal- ime da a ena-
bles dynamic decision-making in p ecision ag icul u e, helping a me s op imize ni ogen
use while minimizing en i onmen al impac s.
2 Ma e ials and me hods
2.1 Da a sou ce
This s udy used da a om a long- e m ield expe imen ocused on win e whea p oduc-
ion a he Dahlem si e in Be lin, Ge many. Ini ia ed in 1984, he expe imen spanned
mul iple c op o a ions un il 1999, p o iding consis en da a collec ion. The si e, cha ac-
e ized by Albic Lu isol soil wi h a loamy sand ex u e, suppo ed a wo- ac o ial design
combining o ganic and mine al ni ogen ea men s. De ails on soil cha ac e is ics, includ-
ing pH, C/N a io, and nu ien con en s (P, K, N), can be ound in [Ko00]. The design
enabled he assessmen o a ying ni ogen a es wi h and wi hou o ganic amendmen s,
aiming o op imize ni ogen applica ion o economic bene i s. The e iliza ion egimes
o c ops, including win e whea , a e summa ized below.
Se ies A: No o ganic e iliza ion. Mine al ni ogen (N) a es: 0 (N0), 150 (N3 o
Po a o), 160 (N3 o Win e Whea ), 120 (N3 o Sp ing Ba ley).
Se ies B: O ganic e iliza ion wi h 300 d /ha manu e o Po a o. Mine al N a es: 0
(N0), 60 (N1), 100 (N2), and 150 (N3) kg/ha o Po a o, 60, 110, and 160 kg/ha o
Win e Whea , and 40, 80, and 120 kg/ha o Sp ing Ba ley.
Se ies C: O ganic e iliza ion includes 60 d /ha S aw + G een Manu e o Po a o,
250 d /ha Bee Lea es o Win e Whea , and 60 d /ha S aw o Sp ing Ba ley.
Mine al N a es as in Se ies B.
The expe imen included de ailed wea he moni o ing, ocusing on key a iables like em-
pe a u e and p ecipi a ion. Fo his s udy, we used a e age empe a u e and o al p ecipi-
a ion du ing he win e whea g owing season (Sep embe o July) om 1986 o 1999,
along wi h annual yield da a ac oss di e en ni ogen ea men plo s.
Da a-d i en ni ogen managemen 323
2.2 Analy ical app oach
This s udy used a model-a e aging app oach o es ima e he a e age EONR o each ea -
men yea , combining mul iple ni ogen esponse models o cap u e a iabili y and en-
hance p edic ion obus ness [MP22; MMP24]. This app oach sui s complex ag icul u al
sys ems wi h non-linea in e ac ions among soil, wea he , and managemen ac o s.
Fou ni ogen esponse models – Mi sche lich, Quad a ic, Quad a ic Pla eau, and Linea
Pla eau – we e indi idually i ed o he yield da a, g ouped by yea and e iliza ion ype
[Ly19]. Due o limi ed mine al N poin s, Se ies A da a was excluded. The Akaike In o -
ma ion C i e ion (AIC) de e mined model quali y, wi h AIC weigh s guiding model con-
ibu ions o he inal a e age EONR. An ANOVA iden i ied signi ican EONR di e -
ences ac oss o ganic e ilize ypes, and ends we e isualized.
To enhance p edic ions, Suppo Vec o Machines (SVM) and Random Fo es (RF) mod-
els analyzed non-linea ela ionships be ween EONR and wea he a iables, ocusing on
o ganic e ilize ypes. SVR models we e s anda dized o balance ea u e con ibu ions,
wi h a adial basis unc ion ke nel se a C=100 and epsilon=0.1 o accu acy. RF models,
which equi ed no s anda diza ion, iden i ied key wea he ac o s a ec ing EONR.
Model pe o mance was assessed ia Roo Mean Squa ed E o (RMSE) and R2 Sco e,
whe e lowe RMSE and highe R2 indica ed be e p edic ion. Combining SVM and RF
models le e aged hei s eng hs o p oduce obus , accu a e EONR p edic ions.
3 Resul s and discussion
3.1 A e age EONR ac oss o ganic a ming ypes
The esul s in Figu e 1 show signi ican di e ences in EONR dynamics be ween he wo
o ganic a ming sys ems. ANOVA esul s indica e a no able di e ence be ween ypes B
and C (p < 0.05), sugges ing ha ni ogen e iciency is in luenced by he a ming sys em,
wi h ype B showing g ea e a iabili y and ype C o e ing mo e s able ni ogen equi e-
men s. Tailo ing ni ogen managemen s a egies o each sys em can imp o e cos -e ec-
i eness and en i onmen al sus ainabili y. Fu u e esea ch could explo e he gene aliza-
bili y o hese indings o o he loca ions and ag oecological condi ions.
324 Cus odio Ma a el and And eas Meye -Au ich
Fig. 1: Compa ison o a e age EONR ac oss o ganic e iliza ion ypes (le ) and ends o a e age
EONR o e he yea s by o ganic e iliza ion ype ( igh )
3.2 Machine lea ning esul s
Machine lea ning model pe o mance a ied signi ican ly by e ilize ype (Fig. 2). Fo
ype B, he Suppo Vec o Reg ession (SVR) model showed a low R2 o 0.07 and high
RMSE (75.47), while o ype C, he R2 was be e a 0.41. These esul s sugges ha SVR
may s uggle wi h high a iabili y in da a [La23; Qi18].
The Random Fo es (RF) models pe o med signi ican ly be e , educing RMSE o 10.95
o ype B and achie ing 1.02 o ype C, e ec i ely cap u ing complex in e ac ions. RF’s
supe io pe o mance aligns wi h indings ha i be e handles a iabili y, making i sui -
able o complex ag icul u al da ase s [La23; Qi18]. This obse a ion ein o ces he end
o using lexible models like RF in p ecision ag icul u e, whe e da ase di e si y demands
obus handling [We22].
The indings highligh he impo ance o selec ing models sui ed o da ase cha ac e is ics
and p edic ion goals. While SVR may sui linea o well-s uc u ed da a, RF o e s a
s onge solu ion o complex ela ionships, enhancing EONR p edic ion accu acy h ough
ad anced machine lea ning.
Da a-d i en ni ogen managemen 325
Fig. 2: Compa ison o Ac ual s. P edic ed EONR Using SVM and Random Fo es Models o O -
ganic Fe ilize Types B and C. The op ow shows he pe o mance o SVM models, while he
bo om ow displays esul s om Random Fo es models
4 Conclusion
This s udy highligh s he po en ial o machine lea ning models, especially Random Fo -
es s, o enhance EONR p edic ion o win e whea . Le e aging his o ical da a on wea he
and ni ogen ea men s, Random Fo es models deli e ed mo e accu a e ni ogen ecom-
menda ions han SVR, pa icula ly o o ganic e ilize ype B. These esul s emphasize
he alue o obus , non-linea models like Random Fo es s, which can handle a iabili y
and suppo p ecise, localized ni ogen managemen . Fu u e esea ch should aim o expand
he da ase and inco po a e addi ional a iables o u he imp o e model accu acy and
applicabili y o sus ainable ag icul u e.
326 Cus odio Ma a el and And eas Meye -Au ich
Bibliog aphy
[La23] de La a, A. e al.: P edic ing si e-speci ic economic op imal ni ogen a e using machine
lea ning me hods and on- a m p ecision expe imen a ion. P ecision Ag icul u e, 24/5,
S. 1792-1812, 2023.
[Ko00] Köhn, W. e al.: Daue düngungs e such (IOSDV) Be lin-Dahlem Deu schland. In:
Kö schens, M. (H sg.): IOSDV: In e na ionale o ganische S icks o daue düngungs e -
suche. Be ich de In e na ionalen A bei sgemeinscha Boden uch ba kei in de In e -
na ionalen Bodenkundlichen Union (IUSS). Bd. 15, LEIPZIG-Halle: UFZ-Umwel o -
schungszen um, 2000.
[LS03] Lo y, J. A.; Scha , P. C.: Yield Goal e sus Del a Yield o P edic ing Fe ilize Ni o-
gen Need in Co n. Ag onomy Jou nal, 95/4, S. 994-999, 2003.
[Ly19] Lyons, S. E. e al.: Ni ogen esponse models o win e ce eals g own o o age. Jou nal
o Ag onomy and C op Science, 205/2, S. 248-261, 2019.
[MMP24] Ma a el, C. E.; Meye -Au ich, A.; Piepho, H.-P.: Model-a e aging as an accu a e ap-
p oach o ex-pos economic op imum ni ogen a e es ima ion. P ecision Ag icul u e,
25/3, S. 1324-1339, 2024.
[MP22] Miguez, F. E.; Po enba ge , H.: How can we es ima e op imum e ilize a es wi h
accu acy and p ecision? Ag icul u al & En i onmen al Le e s, 7/1, e20075, 2022.
[Mo18] Mo is, T. F. e al.: S eng hs and Limi a ions o Ni ogen Ra e Recommenda ions o
Co n and Oppo uni ies o Imp o emen . Ag onomy Jou nal, 110/1, S. 1-37, 2018.
[Qi18] Qin, Z. e al.: Applica ion o Machine Lea ning Me hodologies o P edic ing Co n Eco-
nomic Op imal Ni ogen Ra e. Ag onomy Jou nal, 110/6, S. 2596-2607, 2018.
[Ti02] Tilman, D. e al.: Ag icul u al sus ainabili y and in ensi e p oduc ion p ac ices. Na u e,
418/6898, S. 671-677, 2002.
[We22] Wen, G. e al.: Op imizing machine lea ning-based si e-speci ic ni ogen applica ion
ecommenda ions o canola p oduc ion. Field C ops Resea ch, 288, 108707, 2022.
[Zh15] Zhang, X. e al.: Managing ni ogen o sus ainable de elopmen . Na u e, 528/7580, 51-
59, 2015.
