The use of Artificial Neuronal Networks
to Generate Decision Rules
for Site-Specific Nitrogen Fertilization
Peter Wagner
Martin Luther Universität-Halle, Farm Management Group
www.landw.uni-halle.de/lb/
The problem
??
Almost no decision-making support exists for site specific fertilization,
and no site specific decision rules are available
Foto: Archiv Landtechnik Weihenstephan
Developing decision rules
- Driven by hypotheses testing hypotheses or algorithms in field
experiments
 e.g. - dividing a field into „yield zones“
- transferring „whole field“
decision rules to each sub-field
- NOT driven by hypotheses ex-post analysis of
On-Farm-Research Data
 Data Mining
decision rules have to be tested by „on farm research“
sub-field-spezific
economic evaluation
(cost accounting)
This approach is NOT driven by hypotheses
it is driven by knowledge discovery in databases
What we want to do is extract knowledge (semi-)automatically from
databases. The knowledge must be
• valid
• thus far unknown
• useful
• understandable
Why discover knowledge this way?
„Despite decades of study,
major disagreements and ignorance still exist about
how crop yield responds to managed inputs and non-managed
factors of Production.“ (Bullock and Bullock, 2000, S.96)
Artificial Neural Nets
.
.
.
input
layer
Eingabeschicht
.
.
.
hidden Schicht
layer
verdeckte
.
.
.
output
layer
Ausgabeschicht
The universal useability of ANNs arises because these nets are able
to approximate any arbitrary function.
ANNs are particularly suitable for modeling complex nonlinear
relations which cannot be represented by classic statistical models.
Employing ANNs is a useful approach if there are no theories
available which indicate the concrete model to be used.
Modus Operandi
• For each of the three N-fertilisation-applications an ANN is trained to
predict yield.
• The highest profit generating amount of N-fertilizer is determined via
simulation.
• Decision-tree-algorithms are used to generate decision rules.
• The results are anlysed by means of „response surfaces“.
• Decison rules are tested under real world conditions.
The data (not) needed
In general, there is no need to involve soil survey results (soil
attributes and nutrition state) for yield-prediction.
(Drummond et al. (2000, 2002), Shearer et al. (1999), Liu et al. (2001))
In the eyes of an economist, not using soil survey data is very good,
because it is very expensive to obtain.
What we do need is data of high density that can be collected
automatically (inexpensive and reliable data)!
Data we used (attribute I): yield maps
characteristics:
• indicates soil
properties and yield
potential
• disturbance variables:
outliers, variation of
cutting width, noiseinduced errors etc.
• absolute values vary
over years, in many
regions there are no
stable yield patterns
Data we used (attribute II): apparent soil electrical conductivity (EM 38)
characteristics:
• indicates water holding
capacity
• disturbance variables:
soil etmperature, soil
water status, etc.
• absolute values vary
over years
sending
coil
Foto: Lehrstuhl für Pflanzenernährung TU-München
receiving
coil
Data we used (attribute III): tractor-mounted remote sensing (REIP)
characteristics:
• indicates crop-Nstatus
• meassurement of
reflection in different
wavelengths
• computation into
vegetation index
(REIP)
• disturbance
variables: reflections
of soil, changing light
conditions, etc.
Foto: Lehrstuhl für Pflanzenernährung
data we used (attribute IV): draft force
characteristics:
• indicates top soil
properties
fz
• documented by
the ISO-bus of the
tractor
• disturbance
variables:
oscillation,
compaction of soil,
topography,
acceleration, etc.
Attributes for predicting yield and
generating decision rules
remote sensing
fz
!!
EM 38
yield
prediction of site specific yield
- the basic model EM-38 hist. yield Sensor
X
(mS/m) (dt/ha)
NS
X
fertilizer
(kg/ha)
yield
(dt/ha)
20
64,79,50
?
?
60,70,60
53
45
23
80,77,65
?
?
80,40,90
71
725,3;..
35
40
46,51,60
?
?
90,60,80
64
...
...
...
...
...
...
...
...
REIP
(nm)
d-force
(kN)
SF 0001
723,1;..
21
SF 0002
730,5;..
SF 0003
...
site specific Attributes (x1,x2,..)
site specific yield function:
[non-spatial
attributes
(z1,z2,..)]
variable Input
(FER)
Y = f (x1,x2,..z1,z2,..FER)
the site specific yield function is predicted with an ANN
output (Y)
We must know how yield
reacts to the whole range of
fertilizing possibilities in order
to predict yield based on the
given site-specific information.
Thereforee we need a very
specific experimental design
for training the ANN.
Experimental Farm Görzig
(MLU-Halle)
Field S550, 66 ha,
Winter Wheat (2004)
- Datasource -
Data Audit S550 (homogeous N-application, grid 10m * 10m)
here: relation between yield 2004 and ....
Field
Graph
Type Correlation
Min
Max
Mean Std. Dev Skewness Unique Valid
1
N1
N_1
range
0.123
58.210
64.480
60.039
0.490
5.365
--
660
2
N_2
N_2
range
0.078
30.469
52.032
39.941
1.436
0.376
--
660
3
N3
N_3
range
0.039
30.470
52.030
39.798
1.416
0.560
--
660
REIP32
4 REIP_32
range
0.225
723.410 727.180 725.803
0.580
-0.884
--
660
REIP49
5 REIP_49
range
0.436
725.760 728.990 727.920
0.456
-0.891
--
660
6
range
0.532
21.770
3.841
-0.717
--
660
7 ZUGKRAFT
draft force
range
0.032
2007.420 2339.440 2208.064 48.335
-0.319
--
660
yield_03
8 ERTRAG03
range
0.643
1.588
11.280
6.460
1.528
0.347
--
660
9 ERTRAG04
yield_04
range
--
6.629
10.978
9.411
0.637
-1.118
--
660
range
0.079
120.939 164.062 139.778
2.900
0.322
--
660
range
1.000
711.732 1234.223 1045.463 76.374
-1.113
--
660
10
EM38
EM
38
N_Ges
N_total
NRFC
11 N_Freie_L
39.990
33.916
Available attributes at the
time of training the ANN
decision
variable
point in time:
Ertrag_jj draft force
yield_jj
VS:
EM_38
N_1
yield
Ertrag
decision
variable
EC 32:
Ertrag_jj draft
yield_jj
Zugkraft
force
EM_38
REIP_2
REIP_32
N_1
N_2
yield
Ertrag
decision
variable
EC 49:
Ertrag_jj draft
yield_jj
Zugkraft
force EM_38
soil attributes
REIP_2
REIP_32
REIP_3
REIP_49
in-season attributes
N_1
N_2
N_3
input
Based on the available attributes,
predicting the site-specific yields
was accomplished for each of the three N-applications!
yield
Ertrag
yield (Y)
Site-specific yield prediction
- first N-application T
soil attributes
R
A
histor. 1st N-applications
I
histor. yields (2003)
N
I
ANN
N
G
A
P
P
L
I
C
A
T
I
O
N
histor. training yield (2004)
soil attributes
different N-rates
(decision variable)
expected
yield
Site-specific yield prediction
- second N-application T
soil attributes
histor. in-season attributes I
R
A
histor. 1st N-applications
I
histor. 2nd N-applications
histor. yields (2003)
N
I
ANN
N
G
A
P
P
L
I
C
A
T
I
O
N
histor. training yield (2004)
soil attributes
different N-rates
(decision variable)
current in-season attributes I
expected
yield
Site-specific yield prediction
- third N-application T
soil attributes
histor. in-season attributes I
R
A
histor. 1st N-applications
I
histor. 2nd N-applications
N
histor. 3rd N-applications
histor. in-season attributes II
histor. yields (2003)
I
ANN
N
G
histor. training yield (2004)
A
P
P
L
I
C
A
T
I
O
N
soil attributes
different N-rates
(decision variable)
current in-season attributes I
current in-season
in-season-attributes
I + II
current
attributes II
expectetd
yield
Application of the yield-predicting model
for the economic optimization of N-doses
Beispiel:

SF 01

5.84
7.89
33.0
39.25
70
0.99

yield
NRFCSF*
0
7.32
894.65
10
7.30
897.56
..
...
...
N_2
yield_03 yield_04 d_for EM_38 N_1 REIP_2
* Net-Return over Fertilizer Cost
(per SubField)
prices:
910
WW: 130 €/t
N: 0,65 €/kg
NRFCTF (€)
905
optimum N-rate:
900
100 kg N/ha
895
890
0
20
40
60
80
2. N-rate (kg N/ha)
100
120
Example dataset for genertating decision rules
optimum N-rate
(result of optimization)
yield_03
draft
force
EM_38
N_1
REIP_2
N_2opt
8.1
33
21
40
721.1
30
7.5
27
23
10
723.4
40
8.1
40
22
50
724,3
60
…
…
…
…
…
…
Finally, the knowledge represented in the yield predicting neural
networks is converted into economically-optimized decision rules.
generating decision rules
datasets
(attributes / optimum N-rates)
REIP
EM-38 yield_03 d-force
Nopt
723,1;..
20
64,79,50
?
200
730,5;..
23
80,77,65
?
220
...
...
...
...
...
(x1,x2,..,z1,z2,.. / Nopt);
(…/…); …
decision rules:
e.g. (values assumed)
Data-MiningSystem
(e.g. decision tree)
if REIP > 728,9 and
D-FORCE < 50
and
EM 38 > 23
then
Nopt = 190;
• generated rules should be analyzed by crop production experts
• generated rules have to be tested in field experiments (on farm research)
Response Surface (Görzig S550, 2004, 1st N-application)
(yield_03, training yield 04 )
10.00
9.50
9.00
9.50-10.00
9.00-9.50
8.50
8.50-9.00
8.00-8.50
8.00
7.50-8.00
7.00-7.50
7.50
7.00
Ertrag_04[t/ha]
[t/ha]
yield_04
6.50
100
80
60
N_1 [kg
N/ha]
N_1
[kg/ha]
6.00
40
20
0
3
3.5
4
4.5
5
5.5
6
6.5
7
7.5
Ertrag_03
yield_03[t/ha]
[t/ha]
8
8.5
9
Decision tree (1st
Entscheidungsbaum
zur N-application)
ersten N-Applikation
+
+
<= 41.14
-
yield_03
Ertrag_03
<= 4.42
-
EM_38
<= 25.12
0
2%*
+
+
> 25.12
+
Ertrag03
yield_03
> 41,14
+
> 4.42
<= 7.44
yield_03
Ertrag_03
> 7.44
-
<= 9.20
>9.20
60
50
40
50
40
2%*
80%*
7%*
3%*
6%*
leads
tozuaaeiner
higher
target
N-rate
…führt
höheren
Soll-Menge
leads
to
higher
target
N-rate
…führt
zu einer
gleichbleibenden
no
influence
on
target N-rate
N-rateSoll-Menge
no
influence
on
target
-
EM_38
-
…führt
niedrigeren
Soll-Menge
leads
tozuaaeiner
lower
target N-rate
N-rate
leads
to
lower
target
Ökonomisch optimierte
Soll
Menge
in kg N/ha
economically
optimized
target
N-rates
in kg/ha
* percentage
of sub-fields
* Anteil der Instanzen,
auf die aus dem Trainingssample in diesem
Ast klassifiziert werden
Decision tree (2nd
N-application)
Entscheidungsbaum
zur zweiten
N-Applikation
+
+
<= 725.38
-
REIP_2
<= 725.10
+
-
> 4.88
60
70
+
+
> 725.10
yield_03
Ertrag_03
<= 4.88
REIP_2
+
yield_03
Ertrag_03
-
yield_03
Ertrag_03
<= 7.11
-
<= 6.52
>6.52
60
50
…führt to
zu a
einer
höheren
Soll-Menge
leads
higher
target
N-rate
-
> 725.8
> 7.11
+
REIP_2
<= 725.92
50
> 725.92
50
<= 725.98
30
-
REIP_2
> 725.98
20
Ökonomisch optimierte Soll-Menge in N [kg/ha]
…führt
zu einer on
gleichbleibenden
Soll-Menge
no
influence
target N-rate
-
…führtto
zuaeiner
niedrigeren
leads
lower
target Soll-Menge
N-rate
economically optimized target N-rates in kg/ha
Decision tree (3rd
Entscheidungsbaum
zur N-application)
dritten N-Applikation
-
Ertrag_03
yield_03
<= 7.69
-
<= 725.84
+
+
40
+
…führt to
zu a
einer
höheren
Soll-Menge
leads
higher
target
N-rate
yield_03
REIP_3
50
> 728.54
yield_03
Ertrag_03
<= 6.84
50
-
<= 728.54
-
> 4.49
30
+
> 7.69
yield_03
Ertrag_03
<= 4.49
REIP_2
+
> 725.84
+
> 6.84
70
+
<= 7.57
60
-
Ertrag_03
yield_03
> 7.57
30
Ökonomisch optimierte Soll-Menge in N [kg/ha]
…führt
zu einer on
gleichbleibenden
Soll-Menge
no
influence
target N-rate
-
…führtto
zuaeiner
niedrigeren
leads
lower
target Soll-Menge
N-rate
economically optimized target N-rates in kg/ha
Aim of the experiment: economical
camparison of different fertilizing strategies
(variable rate vs. uniform)
LVG Görzig (MLU-Halle)
S350 (2005) 63,4 ha,
Winter weat (variety:
Compliment):
Betrieb (farm): farmers‘
strategy (uniform
treatment)
Injector: stabilized Nfertilizer, mappingapproach
Sensor: Yara-N-Sensor
Netz (ANN): fertilization
according to ANN decision
rules
N_Var: N-variation
(necessary for the ANN
„learning process“)
N-fertilization with the YARA-N-Sensor®
N-fertilization
with the injector
The strip trail design I
EM_38 values (20m grid)
Average ECa-value (average
of the grids of each variant):
Betrieb (farm): 26,16
Injector: 27,05
Sensor: 26,56
Netz (ANN): 26,88
0.00 - 25.56
25.56 - 26.52
26.52 - 27.58
27.58 - 29.74
The strip trail design II
yield (20m grid)
average yield from 2003
and 2004, in dt/ha
(average of the grids of
each variant):
Betrieb (farm): 84,53
Injector: 84,20
Sensor: 84,60
Netz (ANN): 85,30
0.01 - 79.43
79.44 - 84.10
84.11 - 88.38
88.39 - 98.83
LVG Görzig (MLU-Halle)
S350 (2005) 63,4 ha,
Winter weat:
Betrieb (farm): farmers‘
strategy (uniform treatment)
Injector: stabilized N-fertilizer,
mapping-approach
Sensor: Yara-N-Sensor
Netz (ANN): fertilization
according to ANN decision rules
N_Var: N-variation (necessary
for the ANN „learning process“)
photo: June 28, 2005
variety: Compliment
LVG Görzig (MLU-Halle)
S350 (2005) 63,4 ha,
Winter weat (variety:
Compliment):
Betrieb (farm): farmers‘
strategy (uniform
treatment)
Injector: stabilized Nfertilizer, mappingapproach
Sensor: Yara-N-Sensor
Netz (ANN): fertilization
according to ANN decision
rules
N_Var: N-variation
(necessary for the ANN
„learning process“)
Results S350, Görzig 2005
Unit
„Farm"
"Sensor"
"Injector"
"Netz(ANN)"
N-amount
kg/ha
175
182
190
148
Yield
dt/ha
70,4
73,4
72,1
76,8
Protein content
RP in TS, %
14,3
14,9
15,4
12,9
NRFC1)
€/ha
606
633
615
647
Costs of spreading
fertilizer *
Sensor costs **
Costs for contractor
(injection)
€/ha
15,28
15,28
-
15,28
€/ha
-
4,24
-
4,24
€/ha
-
-
30,4
-
Sum of fertilizing
costs
€/ha
15,28
19,52
30,4
19,52
NRFFC2)
Rank according to
€/ha
NRFFC
591
613
585
628
3.
2. (+22€)
4. (-6€)
1. (+37€)
* Costs for machinery an labor
** Yara-N-Sensor® use for the strategies “Sensor” and “Net”. Assumption: annual use on 1,300 ha, 5 years
depreciation, interest rate 8 % p.a.
1)
Net return over fertilizer cost, calculated with respect to the protein content
2)
Net return over fertilizer and fertilization cost
Results an discussion
• Artificial neural nets seem to offer the possibility of finding useful
decision rules for precision farming.
• The combination of several subfield-specific attributes leads to
significant advantages in modelling yield response as prerequsite for
economic optimization.
• The optimization of each of the N-applications based on the maximum
available Informationen leads to economic advantages.
• The „inverse“ mapping approach to the first N-application seems to be
reasonable.
• (Tractor mounted) remote sensing (REIP) to the 2nd and 3rd
Napplication, in addition to soil attributes, is an important indicator.
• Each field/region needs more or less differnt decision rules.
fin

ICPF_2006_Presentation_Wagner