1. Clay content prediction using on-the-go proximal soil sensor fusion
Salman Tabatabai,
Maria Knadel
Mogens H. Greve
Department of Agroecology
Aarhus University
Denmark

2. Outline
Background
Materials and Methods
Results
Conclusions

3. Background
Materials and Methods
Results
Conclusions

4. The problem High demand for detailed texture maps
Detailed maps require large amount of samples
Collecting this amount of samples is expensive and time-consuming; so is their analysis
Water holding capacity, solute transport, hydraulic conductivity, evaporation, root growth, fertility, and much more is to a large extent influenced by texture parameters.

5. Why on-the-go Sensors? Inexpensive Rapid Highly reproducible
Environmental friendly
Applicable in large scale

6. Why Clay?
Clay maps: Major reference for fertilizer and pesticide application authorization in Denmark
Have the figures for reproducibility of other methods,

7. Background
Materials and Methods
Results
Conclusions

8. Study Sites Silstrup Havris 1-3 Foulum Voulund 1-2 Estrup
Field Name
Area (Ha)
No of data points
Total Cal samples
Soil Type (USDA)
Havris 1
1.5
1928 15
sand
Havris 2
1760
Havris 3
1.9
2061
Silstrup
1.7
2718 20
sandy loam
Estrup
1.2
1801
loamy sand
Foulum 8
4683
Voulund 1 13
7490
Voulund 2 11
5146
Estrup
Map from Google® as retreived in QGIS®

9. On-the-go Soil Sensor Veris® Mobile Sensor Platform:
Vis-NIR Spectrometer
Electrical Conductivity
Temperature
(Veris Technologies, Salina, KS) nm
384 bands
In this slide briefly introduce the sensor
EC_Sh (0-30 cm)
EC_Dp (0-90 cm)

10. Image from Google® as retreived in QGIS®

11. Spectra+EC+T+Pretreatments
Locations
PC Clustering (Spectra)
On-the-go Sensor
Model Input
Calibration samples
Hydrometry (Clay)
R2=0.93
RMSE=1.06
RPIQ=5.2
CV Random
(20 segments)
R2=0.75
RMSE=2.17
RPIQ=2.5
PLS Calibration
CV One-Field-Out (8 segments)
R2=0.93
RMSE=0.96
RPIQ=5.7
Test Set KStone
We have on-the-go sensor towed by a tractor.
Model Input
R2=0.94
RMSE=0.91
RPIQ=6.0
SVM Calibration
CV Random
(20 segments)

12. Locations
Spectra Only
Extensive Noise Removal
PC Clustering (Spectra)
On-the-go Sensor
Model Input
Calibration samples
Hydrometry (Clay)
R2=0.94
RMSE=0.99
RPIQ=5.6
CV Random
(20 segments)
R2=0.74
RMSE=2.25
RPIQ=2.4
PLS Calibration
CV One-Field-Out (8 segments)
R2=0.95
RMSE=0.79
RPIQ=7.0
Test Set KStone
We have on-the-go sensor towed by a tractor.
Model Input
R2=0.94
RMSE=0.95
RPIQ=5.8
SVM Calibration
CV Random
(20 segments)

13. Background
Materials and Methods
Results
Conclusions

14. Although the two modeling approaches overall similar results, RMSEV in the spectra only models are lower than the fused sensor models and the R2 are higher.
In the case of Kstoen Test Set, the cleaned spectra RMSEV is about 18 percent lower compared to the fused sensors.
In other cases, there are negligable differences in R2 and RMSEV between the two model ing approaches.
RPIQ=7.0
4 PLS Factors

15. Background
Materials and Methods
Results
Conclusions

16. Conclusions
One-the-go sensors can predict clay content accurately and robustly even in fields with high variability of clay content and in different conditions
In cases that no pretreatments on spectra is possible, fusing spectral data with EC and T may improve the predictions
Meticulous and careful preprocessing and noise removal of spectral data can provide the best predictions without the need for fusing it with any auxiliary data

17. Aarhus University, Denmark
Conclusions (cont’d)
It is imperative to note that the findings of this research might be only applicable in areas with sand as major soil particle size (e.g. in Denmark)
More research on soils with higher clay and silt content is required before a conclusive decision is made on designing new mobile sensor platforms.
Additive note:
There is a need for further development of on-the-go sensor spectral library in order to enhance the model usability in highly different fields.
Salman Tabatabai
Aarhus University, Denmark