NJF presentation - Estimating the content of clover and

ESTIMATING THE CONTENT OF
CLOVER AND GRASS IN THE
SWARD USING A CONSUMER
CAMERA AND IMAGE
PROCESSING
A. K. Mortensen1, H. Karstoft1, K. Søegaard2 and R. N. Jørgensen1
1Department
of Engineering, Aarhus University
2Department of Agroecology, Aarhus University
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
OUTLINE

Motivation

Test bed and data acquisition

Methodology

Results

Conclusion and future work
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
MOTIVATION

Clover and grass is sown as a catch crop and as a feed crop

However, it is unknown how the distribution is in the field

The distribution and total dry matter determines
› the N-uptake in the field
› As feed for dairy cows
› intake
› milk yield

Current method is destructive analysis
› Labour intensive
› Cumbersome
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
TEST BED AND DATA ACQUISITION

Plot experiment (2013) near Research Center Foulum, Denmark

Two different seed mixtures:
› “Blanding 35”: perennial ryegrass and white clover
› “Blanding 45”: perennial ryegrass, festulolium, red clover and white clover

Photographed and cut on 4 different occasions
› Cuts made within 3 days after photograph
› Photographed from above
› 45 images
› Dry matter analysis in laboratory
› Dry matter: 195 kg/ha  6111 kg/ha
› Clover: 10-72%
› Grass: 26-90%
› Weed: 0-3%
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
TEST BED AND DATA ACQUISITION
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
TEST BED AND DATA ACQUISITION
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
METHODOLOGY
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
METHODOLOGY

Illumination classification

Direct and indirect sun light

Histogram of pixel intensities

Cross correlation used for
classification
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
METHODOLOGY

Coverage estimation

Remove background:
› Soil, dead plant
material and deep
shadows

Extract clover leafs:
› Inverted edge image +
erosion

Grass:
› Remaining

Trained on patches
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
METHODOLOGY

Transformation of coverage to dry matter distribution
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
RESULTS

Coverage estimation
Direct light
AU
Indirect light
Sh*
So*
G*
C*
F*
W*
∑
Sh
66%
1%
25%
8%
0%
0%
2955
So
0%
0%
0%
0%
0%
0%
G
9%
0%
73%
17%
0%
C
13%
0%
39%
48%
F
30%
20%
46%
W
24%
0%
∑*
3361
98
AARHUS
UNIVERSITY
Sh*
So*
G*
C*
F*
W*
∑
Sh
52%
11%
29%
9%
0%
0%
3934
0
So
11%
87%
2%
1%
0%
0%
2422
0%
4745
G
11%
1%
65%
23%
0%
0%
6757
0%
0%
6565
C
3%
0%
29%
68%
0%
0%
13491
4%
0%
0%
254
F
18%
0%
77%
5%
0%
0%
79
71%
4%
0%
0%
181
W
7%
0%
52%
41%
0%
0%
617
6986
4255
0
0
-
∑*
3541
2654
9795
11310
0
0
-
NJF
23 NOVEMBER 2014
RESULTS

Dry matter distribution
Mean
Test set:
Clover
Grass
Training set:
Clover
Grass

Error (%-points)
Std.
Max
Min
Absolute error (%-points)
Mean
Std.
Max
Min
-2.1
2.6
9.8
10.4
19.1
17.4
-15.6
-18.7
7.9
8.8
5.8
5.7
19.1
18.7
0.0
1.3
0.0
0.0
12.9
13.5
18.9
51.0
-50.3
-20.7
8.6
9.2
9.4
9.7
50.3
51.0
0.2
0.2
No correlation between error and mixture, clover dry matter or total dry matter.
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
CONCLUSION AND FUTURE WORK

It is possible with a reasonable accuracy (8-9%-points)

Greatest source of error
› Coverage estimation

Room for improvement
› Better estimation of coverage
› Texture analysis
› Illumination invariant model
› Include growth models
› Time since last harvest
› Temperature sum
› Available water
AU
AARHUS
UNIVERSITY
NJF
23 NOVEMBER 2014
AU
AARHUS
UNIVERSITY