=Paper=
{{Paper
|id=Vol-2030/HAICTA_2017_paper80
|storemode=property
|title=Applying Different Remote Sensing Data to Determine Relative Biomass Estimations of Cereals for Precision Fertilization Task Generation
|pdfUrl=https://ceur-ws.org/Vol-2030/HAICTA_2017_paper80.pdf
|volume=Vol-2030
|authors=Jere Kaivosoja,Roope Näsi,Teemu Hakala,Niko Viljanen,Eija Honkavaara
|dblpUrl=https://dblp.org/rec/conf/haicta/KaivosojaNHVH17
}}
==Applying Different Remote Sensing Data to Determine Relative Biomass Estimations of Cereals for Precision Fertilization Task Generation==
https://ceur-ws.org/Vol-2030/HAICTA_2017_paper80.pdf
Applying Different Remote Sensing Data to Determine
Relative Biomass Estimations of Cereals for Precision
Fertilization Task Generation
Jere Kaivosoja1, Roope Näsi2, Teemu Hakala2, Niko Viljanen2, Eija Honkavaara2
1
Green Technology Unit, Natural Resources Institute Finland (LUKE), Vakolantie 55, 03400
Vihti, Finland, e-mail: jere.kaivosoja@luke.fi
2
Finnish Geospatial Research Institute, Finland, e-mail: eija.honkavaara@nls.fi
Abstract. Recently, the area of passive remote sensing of agricultural fields
has been developing fast. The prices of RPAS (remotely piloted aircraft
system) equipment has gone down and new suitable sensors are coming into
markets while simultaneously new and free relevant satellite data has become
available. One of the most used applications for these methodologies is to
calculate the relative biomass as a basis for additional nitrogen fertilization. In
this work, we study the difference of biomass estimations based on Sentinel-2
imagery, tractor implemented commercial measurement system, a low-cost
RPAS equipment with commercial software and a hyperspectral imaging
system implemented in a professional RPAS system in fertilization planning.
Our study revealed that while there was a 23 % spatial variation in our test
field’s yield, the relative biomass estimations for fertilization planning during
the growing season varied 22 % on average although they were visually very
alike.
Keywords: Sentinel-2, RPAS, variable rate application (VRA), fertilization
1 Introduction
The core idea of precision farming is to spatially and timely optimize the farming
inputs to maximize the farming outcomes while reducing the environmental stress.
Nitrogen fertilizers are one of the core inputs in plant production. An insufficient
dosage of the nitrogen fertilizer for cereal crops can decrease the yield and quality of
the yield. Excess of nitrogen causes a risk of a flattening of the growth causing yield
losses. Also unused nitrogen in the soil leaches to the environment throughout the
growing period and after.
Already developed precision nitrogen application methods for crops utilize an
optical sensing of the growth status during the growing season to determine how
much additional nitrogen is needed in different areas of a field. Sensing may take
place from satellites, aircrafts, RPAS’s (remotely piloted aircraft system), working
machinery or handheld devices. Recently there has been a fast development in this
area of passive remote sensing and the productization is in progress. The prices of
670
RPAS equipment have gone down and new sensor technologies are coming into
markets. Also new, free and relevant satellite technology has become available for
the environmental mapping. In Europe, the new Sentinel-2 satellites are providing
useful data several times per week with up to 10 x 10 meter accuracy.
One of the most common applications for these methodologies in agriculture is to
calculate relative biomass as a basis for additional nitrogen fertilization. Typically,
these sensing systems compare a red and near infrared wavelengths by measuring a
normalized difference vegetation index (NDVI) or its variants. Then an implemented
decision support system (DSS) produces an estimate for the required nitrogen
fertilization need. This DSS system requires calibration information about crop’s
remaining nitrogen needs and responsiveness according to the predicted yield
potential (Raun et al., 2005, Lukina et al., 2001) being important factor for the
nitrogen fertilization. Data from other sources are usually combined with remote
sensing as inputs to decision support systems for determining nitrogen application
rates (Shanahan et al. 2008; van Evert et al. 2012, Kaivosoja et al., 2013).
Hyperspectral imaging for example was found to be a promising method for
agricultural purposes (Bareth et al. 2015) and obtaining separate biomass and
nitrogen content (Honkavaara et al. 2013, Pölönen et al. 2013) for additional
fertilization need determination.
In practice, the basic NDVI maps indicate the amount of green mass in the field.
However, the method is not able to differentiate situations of a low growth density
with high nitrogen content from those of high growth density and a low nitrogen
content. Thus, generating nitrogen fertilization plans based only on NDVI map might
not be the best solution in all of the cases so many supporting optical methodologies
has been developed. Pena-Yewtukhiw et al. (2015) found out that even the sensor
output difference of 0.05 NDVI units could strongly affect the resulting nitrogen rate
prescription, depending on the selected algorithms. Also image mosaics that are
mandatory with RPAS sensing may create large radiometric errors that effect on
spectral vegetation indices (Rasmussen et al. 2016).
Dong et al. (2015) presented 28 chlorophyll-related vegetation indices suitable to
be applied with Sentinel-2 data and by simulation studies; they found out that
incorporating red-edge reflectance (around 700nm) improved the estimates for
assessing vegetation growth rate and predicting crop productivity. Also, Hunt et al.
(2017) noted that assessing red-edge detection could make a difference in
determining nitrogen applications to potato. In their study, they did not found RPAS
beneficial to the WorldView-2 satellite data. The case is similar with the Sentinel-2
imagery in Europe since the resolution of the red edge is coarser.
That is also what current commercial solutions support. The tractor implemented
YARA N-Sensor five spectrometer detects the wavelengths of 550nm, 650nm,
700nm, 710nm and 840nm (Varco, 2010). The gained economic benefits of this
tractor implemented solution have been around 5 % (Nissen, 2012). Typically drone
installed Parrot Sequoia multispectral camera measures the wavelengths of RGB,
550nm, 660nm, 735nm, 790nm. The most accurate wavelengths of Sentinel-2
satellite are 490nm, 560nm, 665nm, 842nm with 10 meter spatial resolution and
705nm, 740nm, 783nm, 865, 1610nm, 2190nm with 20 meter spatial resolution. In
Finland, the average field size is less than 4 hectares which makes it difficult to
exploit coarser data efficiently.
671
Many new technologies are coming available, but since the productization for
agricultural purposes is continuously developing, the farmers are somewhat left alone
on how to really apply them and how to get the best benefits out of them and which
methods would be the most suitable for their purpose.
This paper has three research questions: 1) how much there is typically variation
in a Finnish field, meaning that how much we should typically adjust the amount of
fertilizers? 2) How much there is variation of relative biomass estimations based on
different remote sensing data obtained for the same purpose? 3) What is the effect of
the determined variations in contrast to experimental but logical precision
fertilization application variations? The main goal is to demonstrate in real
conditions, how much difference there are in biomass estimations in contrast to
actual fertilization task variations.
2 Material and methods
The test area was about 20 ha cereal crop field in southern Finland in Vihti, sowed at
29 May 2016. The overview picture of the field during the 2016 crowing season is
presented in a Fig. 1. The field was evenly treated although a 12 meter wide not
treated stripe was left in the middle of the field to have a bare soil reference.
Fig. 1. A slant view of the test field showing the high biomass area in the left size and the not
seeded stripe in the middle, (date 16.7.)
First, to have a concrete knowledge about variations in our field, we analyzed our
combine harvester data to measure the yield variation in the selected test field and in
the fields nearby. We analyzed the yield data of barley and wheat from the years
2015, 2014, 2013. In total, 20 harvestings with an average field plot size of 6.4 ha.
Those fields were evenly threated (no precision farming) and the harvesting was
done with Sampo Comia C4 combine harvester with Ceres 8000 yield monitor,
which logged position and filtered yield data with 5 Hz interval. We filtered out less
than 900kg/ha measurements and exceptionally high yield values from the yield data.
672
Next, we calculated a variance for each harvesting operation. To study the effect
of combine harvester measurement system effect, we added a separate moving
average of five for the logged data.
Next we studied different remote sensing data. Fig. 2. presents relative biomass
maps based on different remote sensing technologies: a professional UAV
(unmanned aerial vehicle) with FPI (Fabry Perot Interferometer)-hyperspectral
camera, consumer level Phantom 4 UAV with RGB-camera, a tractor implemented
Yara N-sensor and Sentinel-2 satellite image. More detailed descriptions of data
processing of these data are presented by Näsi et al. (2017). These maps represented
the starting point of this work. The middle part in the N-Sensor map (Fig. 2.) was not
measured due to low amount of biomass.
Fig. 2. Relative biomass estimations based on professional UAV with FPI-camera, consumer
level Phantom 4 drone with RGB-camera, tractor implemented Yara N-sensor and Sentinel-2
satellite image.
Our next step was to use the different source data to produce precision nitrogen
fertilization tasks without additional data. We used farmer knowledge to heuristically
adjust the tasks in a similar manner. All remote sensing data that was used with our
calculations are presented in the following list, including a name of data,
measurement instrument and platform, classification type and imaging date in 2016.
• Tractor: Yara-N-sensor measurements, internal classification, driving, date
16.7.
• Satellite1: Sentinel-2 satellite, NDVI classification, image date 2.6.
• Satellite2: Sentinel-2 satellite, NDVI classification, image date 9.7.
• proUAV FPI (Fabry-Pérot interferometer) hyperspectral camera, NDVI
classification, imaging date 4.7.
• rgbUAV1: Phantom 4, stock RGB camera, classification VARI Visible
Atmospherically Resistant Index (G-R)/(G+R-B) (Gitelson et. al 2002) with
DroneDeploy software, 16.7.
• rgbUAV1task: rgbUAV1 data classified with DroneDeploy, 16.7.
• rgbUAV2: 16.7. Phantom 4, stock RGB camera, VARI classification with
DroneDeploy software, 16.7., constantly changing cloud cover
• Yield Map: Yield map based on combine harvester point data and surface
fitting by using inverse distance weighting (5 m circle search distance and
weighting power of 1), 23.9.
673
All the different tasks were planned in order to have three different fertilization
levels: 20 kg/ha, 30 kg/ha and 40 kg/ha according to farmer’s understanding of the
additional fertilization need. Three levels were selected to be practical with present
farm machinery: there are always some delays and inaccuracies with precision
adjustments, so it is not practical to adjust the machinery continuously. These
fertilization tasks were calculated as 1 m grid maps. Then we compared these maps
to accurately located biomass samples taken during the growing season (Näsi et al.
2017). We used 18 samples out of total 36 samples having the most homogenous
surroundings around them. Then we also calculated the correlation between biomass
amount and nitrogen content from the vegetation samples to see their correlation.
To evaluate the effect of the usage of other data sources, we demonstrated possible
task variations by applying previous yield maps, vegetation samples, farmer’s know-
how and commercial software for data. We selected four different cases which were
as follows:
1. Previous yield maps and Sentinel-2 data 9.7. First, we evenly balanced and then
summed three consecutive yield maps. Next we scaled the final map values by
using farmer’s heuristic knowledge. The parameters were: min 0.4, Max 1.6,
Mean 1.07, Std. deviation 0.19. Then we used this to multiply the Sentinel-2
NDVI-map. Then we applied contouring method to generate four application
rate levels, and finally farmer decided the actual fertilization amounts.
2. Consumer UAV with RGB-camera (Phantom 4), Dronedeploy vegetation
classification (VARI) and farmer estimates for actual fertilization amounts.
3. NDVI classification from professional UAV with FPI camera. We used
supervised K-means teaching based on vegetation samples (nitrogen content)
including 36 samples from all around the field (Näsi et al. 2017) and categorized
into four classes by the farmer (none-now-med-high. We used it to supervise
proUAV data to four classes. Then actual fertilization amounts were decided
according to farmer’s knowledge.
4. Consumer UAV (Phantom 4) with RGB-camera, VARI calculated with
DroneDeploy software, added with farmer teaching (polygons drawn by the
farmer including wanted nitrogen input) by using K-means methodology. In this
study, the farmer drew the wanted fertilization amounts on top of a plain RGB-
map. Then these areas were used to teach the VARI raster map. Finally, the
contouring method was applied as in other cases.
We used 0-10-20-30 kg/ha fertilization steps for these data, because these other
data suggested lower fertilization rates and it would not have be reasonable to
compare these with the first classification results.
674
3 Results
The average of the yield amount variances in our fields was 32.7 %. Our test field
yield had a variance of 23.3 %, the histogram is presented in Fig. 3. By applying the
moving average, the variance was lowered only by 0.5 percentage points. This is
indicating at least a 30 % variance in yields on average in our test fields in Finland.
The total yields of our fields were 4.6 t/ha on average and the average variance was
1.8 t/ha. During the summer 2016, our test field had an exceptionally low yield on
average.
Fig. 3. Test fields yield histogram (yield amount and number of measurement points)
The different fertilization tasks based only on the remote sensing data are
presented in a Fig. 4. Together with a relevant yield map that was harvested more
than two months later. Table 1. compares these different maps by showing the
average difference between application rates: if A=20kg/ha and B=30kg/ha, A
compared with B is 10/20=0.5 different and B compared with A is 10/30=0.33
different. On average, the difference between calculated application rates was 22%.
675
Fig. 4. Fertilization tasks based on remote sensing data and the final relative yield map
676
Table 1. Difference between calculated application rates
rgb2 Sat2 Tractor proU rgb1t rgb1 Sat1
rgbUAV2 0.00 0.17 0.41 0.18 0.21 0.22 0.29
Satellite2 0.18 0.00 0.39 0.20 0.21 0.19 0.26
Tractor 0.30 0.28 0.00 0.25 0.25 0.23 0.21
proUAV 0.17 0.19 0.33 0.00 0.17 0.17 0.28
rgbUAV1task 0.19 0.19 0.30 0.16 0.00 0.17 0.19
rgbUAV1 0.19 0.17 0.28 0.16 0.17 0.00 0.21
Satellite1 0.24 0.21 0.23 0.24 0.17 0.19 0.00
The correlations between biomass samples (18 spots) and calculated application
rates with different methods are presented in Fig. 5. The tractor data had the highest
correlation of 0.63. The correlation between biomass and nitrogen based on
vegetation samples is also presented being -0.19.
Fig. 5. Correlation between vegetation sample biomass and determined application rates
Next we present the demonstrative task maps, which combined other data to
remote sensing, with the following early presented methodologies:
1. Previous yield maps and Sentinel-2 data
2. Consumer UAV with RGB-camera, Dronedeploy classification and
farmer’s nitrogen level estimates
3. NDVI classification from professional UAV with FPI camera, teaching
with vegetation samples
4. Consumer UAV with RGB-camera, RGVI-index and farmer teaching
677
The nitrogen fertilization tasks were clearly deviating. The following Fig. 6.
illustrates the different generated tasks.
Fig. 6. Different fertilization tasks based on remote sensing and external data
4 Discussion and conclusions
Our test fields had the 30% yield variation in average so there is huge potential in
precision farming activities. All the tested remote sensing methods managed to
estimate the relative differences of biomass. The optimal timing for the additional
fertilization would have been in the middle of July, and even the Sentinel-2 NDVI-
map in early June estimated visually correctly the relative biomass. However, when
developed into precision fertilization tasks, the relative biomass estimations produced
a 22 % variation in an average, when the planned fertilization was 20kg/ha - 30kg/ha
- 40kg/ha in all the cases, including the aim to produce similar looking maps.
Similarly, Pena-Yewtukhiw et al. (2015) stated that even a slight difference in the
single task generation parameter could produce a remarkable difference in the end.
Also in our measurements, the correlations to the biomass samples were low.
When other parameters were used for task generations, the differences were large
even when based on visual estimations. The main difference between images 2 and 4
in Fig. 6 is that in image 4, the farmer decided the effective area for the fertilizer,
while in the image 2 the area was decided by the RGVI difference. The optimistic
attitude of the farmer can be seen as the application rate is higher in the image 2
(Fig.4.).
The visual study of Fig. 4. shows that Sentinel-2 data from 2.6. and 9.7. are giving
very similar information. This is indicating that these NDVI-level differences can be
spotted even in a very early stage of growth.
The N-sensor values in Table 1. Were lower than others and were suggesting less
fertilization. This was true according to our true vegetation samples and the yield
map, there were mostly enough nitrogen resources for the plant. The N-sensor data
(Tractor) had the highest correlation to the biomass samples. So without concrete
relations, the RPAS and satellite data were exaggerating the nitrogen need.
When the hyperspectral imagery was used only for the biomass estimations as we
did, there were no significant advantages seen. We assume that the usage of a
678
multispectral camera would have similar results here. In both cases, additional
estimations such as vegetation nitrogen content estimations would be essential.
As the main conclusion of this work, there is a large variation within cereal fields
in Finland, the relative difference was easy to determine with different remote
sensing methods, but there is huge step needed to use these biomass variations in a
consistent way. Just picking up a drone or a free satellite image would possibly not
give a sufficient knowledge for additional fertilization.
We should also note that the yield of our test field was low and the areal
differences between crops were very similar during the entire growing season. These
factors can be very different in different years when there is for example lack of
water, so the very generalizing conclusions of the goodness of the relative biomass
estimations cannot be drawn.
Acknowledgments. We acknowledge ESA (ESRIN/Contract No. 4000117401/16/I-
NB), Tekes (1617/31/2016) and ICT-AGRI ERANET project Geowebagri II for
funding the project. We are grateful to Paula Litkey and Milos Pandžića for their
support in preprocessing of the Sentinel images.
References
1. Bareth G., Aasen H., Bendig J., Gnyp M. L., Bolten A., Jung A., Michels R.,
Soukkamäki J. (2015) Low-weight and UAV-based hyperspectral full-frame
cameras for monitoring crops: spectral comparison with portable
spectroradiometer measurements. 2015. Photogrammetrie - Fernerkundung -
Geoinformation PFG 2015 (1), pp.69-79. doi: 10.1127/pfg/2015/0256.
2. Dong, T., Meng, J., Shang, J., Liu, J. and Wu, B. (2015) Evaluation of
Chlorophyll-Related Vegetation Indices Using Simulated Sentinel-2 Data for
Estimation of Crop Fraction of Absorbed Photosynthetically Active Radiation,
IEEE Journal of Selected Topics in Applied Earth Observations and Remote
Sensing, 8:8:2015.
3. Gitelson, A. A., Kaufman, Y., Stark, R., Rundquist, D., (2002) Novel algorithms
for remote estimation of vegetation fraction, Remote Sens. Environ., 80:76-87.
4. Honkavaara, E., Saari, H., Kaivosoja, J., Pölönen, I., Hakala, T., Litkey, P.,
Mäkynen, J., and Pesonen, L., (2013) Processing and Assessment of
Spectrometric, Stereoscopic Imagery Collected Using a Lightweight UAV
Spectral Camera for Precision Agriculture. Remote Sensing, 5(10), pp. 5006-
5039
5. Hunt E., Horneck, D., Spinelli, C., Turner, R., Bruce, A., Gadler, D., Brungardt,
J., Hamm, P., (2017) Monitoring Nitrogen Status of Potatoes Using Small
Unmanned Aerial Vehicles, Precision Agric, 1-20, doi:10.1007/s11119-017-
9518-5.
679
6. Kaivosoja J., Pesonen L., Kleemola J., Pölönen I., Salo H., Honkavaara E., Saari
H., Mäkynen J., Rajala A. (2013) A case study of a precision fertilizer application
task generation for wheat based on classified hyperspectral data from UAV
combined with farm history data, Proceedings of SPIE 2013 Vol. 8887, 88870H.
7. Lukina, E., Freeman, K., Wynn, K., Thompson, W., Mullen, R., Stone M, et al.
(2001) Nitrogen fertilization optimization algorithm based on in-season estimates
of yield and plant nitrogenuptake. J. Plant Nutr. 24:885–898. doi:10.1081/PLN-
100103780.
8. Marti, J., Bort, J., Slafer, G.A. and Araus, J.L. (2007) Can wheat yield be
assessed by early measurements of Normalized Difference Vegetation Index?.
Annals of Applied Biology, 150: 253–257. doi:10.1111/j.1744-
7348.2007.00126.x.
9. Nissen K. (2012) Yara N-Sensor – sensible sensing, Testing and certification of
agricultural machinery, Riga, Latvia 16 - 18 October 2012 / Nils Bjugstad, Eskil
Nilsson and Gints Birzietis (Eds.). NJF Report 8 6: 69-70. (2012).
10. Näsi, R., Viljanen, N., Kaivosoja, J., Hakala, T., Pandžić, M., Markelin, L.,
Honkavaara, E. (2017) Assesment of various remote sensing technologies in
biomass and nitrogen content estimations using an agricultural test field. Remote
Sensing (in press).
11. Pena-Yewtukhiw, E., Grove, J., Schwab, G., (2015) Fertilizer Nitrogen Rate
Prescription, Interpretational Algorithms, and Individual Sensor Performance in
an Array. Agron. J. 107:1691-1700. doi:10.2134/agronj14.0573.
12. Pölönen I., Saari H., Kaivosoja J., Honkavaara E., Pesonen, L. (2013)
Hyperspectral imaging based biomass and nitrogen content estimations from
light-weight UAV. Proceedings of SPIE 2013 Vol. 8887, 88870J
13. Rasmussen, J., Ntakos, G., Nielson, J., Svensgaard, J., Poulsen, R. N., &
Christensen, S. (2016) Are vegetation indices derived from consumer-grade
cameras mounted on UAVs sufficiently reliable for assessing experimental plots?
European Journal of Agronomy, 74, 75–92.
14. Raun, W., Solie, J., Stone, M., Martin, K., Freeman, K., Mullen, R., Zhang, H.,
Schepers, J., Johnson, G. (2005) Optical sensor based algorithm for crop nitrogen
fertilization. Commun. Soil Sci. Plant Anal. 36:2759–2781.
doi:10.1080/00103620500303988.
15. Shanahan, J., Kitchen, N., Raun, W., Schepers, J., (2008) Responsive in-season
nitrogen management for cereals. Computers and Electronics in Agriculture, 61,
51–62.
16. van Evert, F. Booij, R., Jukema, J., Berge, H., Uenk, D., Meurs, E., et al. (2012)
Using crop reflectance to determine sidedress N rate in potato saves N and
maintains yield. European Journal of Agronomy, 43, 58–67.
17. Varco J. (2010) Sensor Based Fertilizer Nitrogen Management. Crop
Management Seminar. Nov. 9-11, 2010 Memphis, TN, USA. Available online:
http://www.cottoninc.com/fiber/Agricultural-Research/Agricultural-Meetings-
Conferences/Crop-Management-Seminars/2010%20Presentations/23-PA-Sensor-
Based-Fertilizer-Nitrogen-Management.pdf (Assessed on 15.5.2017).
680