Basic Civil Engineering notes on Transportation Engineering & Modes of Transport
Can small NIR calibrations at the farm scale be an alternative to large calibrations for farm soil mapping of soil type?
1. Bo Stenberg and Johanna Wetterlind
Swedish University of Agricultural Sciences
Rikard Westbom and Daniel Olsson
Eurofins Food & Agro Testing Sweden AB
Can small NIR calibrations at the
farm scale be an alternative to
large calibrations for farm soil
mapping of soil type?
2. NIR for farm soil mapping of soil
organic matter and clay content
Standard procedure today:
Top soil sampling – one sample per ha
Plant available nutrients and pH analyzed on
every sample
Texture and SOM on every third sample – if
at all
3. Our mission:
SOM on every sample with NIR
Texture on every sample with NIR
Plant available nutrients and pH also with NIR
Evaluate small farm scale calibrations for
implementation in a commercial lab
How small can calibrations be?
Keep it simple!
Clay and SOM presented here
5. Seven farm sites in Sweden – 41-206 ha
picked from the production line at Eurofins
1 sample per ha
6. Methods:
Samples – air dried and sieved to 2 mm
SOM – loss of ignition corrected for structural
water
Clay – sedimentation (<2µ)
NIR – 1300-2500nm, NIRSystem 5000
(FOSS) – the system in use at the Eurofins
lab
Everything analyzed on all samples
7. Calibratons:
First derivative spectral pre-treatment
Farm-wise PLS
1. Every third for calibration to predict the
remaining
2. Kennard-Stone selected calibration samples
in steps of 10 up to N/2 to predict the
remaining
10. Outlier detection
Important – especially in small calibrations
• Outliers in the calibration step – Manageble
• Outliers in the prediciton step – Not preferable
• Most methods removed too many samples
• Calibration step – X-Y relationship if CV bad (u-t
scores)
• Prediction step leverage >1 for prediction samples
16. Relationship between SD and RMSE of 26
published large scale calibrations – SOM
Standard Deviation (% SOM)
0 1 2 3 4 5 6 7
RMSE(%SOM)
0.0
0.5
1.0
1.5
2.0
2.5
RMSE=0.20+0.31*SD
r2
=0.85
17. Standard Deviation (% SOM)
0 1 2 3 4 5 6 7
RMSE(%SOM)
0.0
0.5
1.0
1.5
2.0
2.5
RMSE=0.20+0.31*SD
r2
=0.85
Some published field and farm scale pre-
dictions in relation to the large scale
regressionline – SOM
18. Standard Deviation (% SOM)
0 1 2 3 4 5 6 7
RMSE(%SOM)
0.0
0.5
1.0
1.5
2.0
2.5
RMSE=0.20+0.31*SD
r2
=0.85
The present farm scale pre-dictions in relation
to the large scale regressionline – SOM
19. Standard deviation (%)
0 5 10 15 20 25 30
RMSE(%)
0
2
4
6
8
10
12
14
RMSE = -0.02 + 0.42 * SD
r ² = 0.77
The present farm scale predictions in relation
to the large scale regressionline and
calibrations – Clay
RPD (SD/RMSE) is
the same along the
line (~2.3)
Are the often used
RPD thresholds
relevant?
20. Lime requirement at Askersund (206 ha)
1/ha 1/3 ha
1/ha predicted
from KS40
About 9 tons ha-1
on average
21. Diff. all – 1/3 ha Diff. all – pred.
from KS40
Lime requirement at Askersund (206 ha)
7% 4%
22. Lime requirement at Västerås (175 ha)
1/ha 1/3 ha
1/ha predicted
from KS40
About 16 tons
ha-1 on average
23. Lime requirement at Västerås (175 ha)
Diff. all –
1/3 ha
Diff. all – pred.
from KS40
16% 2%
24. Conclusions
• It works – it seems
• Worthwhile for the labs to consider
• Some more thinking about outlier detection
• Calculations on the economy needs to be done