Previous PageTable Of ContentsNext Page

An index for evaluating crop production variability from remote and proximal sensor data

Ronaldo de Oliveira1 and Brett Whelan2

1 Embrapa Solos – National Center for Soil Research, Rio de Janeiro, RJ, Brazil, 20460-000. Email r.oliveira@usyd.edu.au
2 Australian Centre for Precision Agriculture, University of Sydney NSW 2006.

Abstract

A key task in the process of developing decision support for differential crop management is to effectively quantify and rank the degree of inherent, manageable production variability within a field. It would be ideal if such assessment could be undertaken using remote or proximally sensed data that could be gathered external to farm operations. This paper investigates the application of a parametric opportunity index (Yieldex) previously developed for use with yield monitor data, to crop reflectance imagery and apparent soil electrical conductivity data (ECa). The method is based on variogram analysis of the data and is used to quantify the degree of within-field variability as a function of the overall magnitude of variation and the cohesion of spatial variability patterns relative to the present ability of variable-rate machinery to react. Results obtained from ECa and imagery data show that both may be suitable for use in an ‘opportunity index’. Soil ECa data from root zone depth appears to provide a more suitable general index than imagery data gathered mid-season in broad acre crops. Indices calculated from imagery should prove useful in single-season assessments following good growing conditions.

Key Words

precision agriculture, variability, production, index, management

Introduction

The “null hypothesis” of Precision Agriculture (PA) suggests that traditional, uniformly applied management practices may be optimal where a lack of well structured spatial variability is observed (Whelan and McBratney, 2000). Although strategic decision-tree concepts linked to testing the “null hypothesis” have now been described down to the detailed level of individual field-operation information requirements (Fountas et. al., 2006), limited attention has been given to the formulation and validation of quantitative methods to simultaneously evaluate the magnitude and pattern of production variability encountered on-farm.

In fact, relevant analysis of the scale of within-field variability of crop yield has been mostly restricted to the analysis of numerical statistics (Gangloff et. al., 2004; Reyniers et. al., 2006), or in conjunction with visual interpretation of interpolated maps (Taylor et. al., 2003). These analyses are often tailored to individual experiments where proximal and remote sensors are used to explore the relationship between crop influencing factors. Often such processes generalise continuous, dense numerical data sets into a discrete categorical yield classification of high-medium-low on a site-by-site basis. The spatial structure, its relationship to the operational specifications of the machinery that farmers use for input management, and the degree of variation relative to a critical benchmark is neglected.

It could be argued that the lack of development of a more holistic approach to quantifying within-field variability as an indicator of the potential for site-specific crop management (SSCM) may be delaying the broad adoption of PA (McBratney et. al., 2005). Indeed, problems related to the development of decision support systems (DSS) for PA have long prevailed over the expectations within the agricultural community that the application of computer technology would revolutionise management of the contemporary family farm (McCown, 2002).

Following on from the work of Pringle et al., (2003), de Oliveira et al., (2007) explored assessing the opportunity for adoption of PA based on a quantitative index that is a function of both the magnitude and spatial structure of the observed yield variation. The procedure has proven robust in fields with both stationary and non-stationary variation in crop yield as recorded using yield monitors at harvest. However, the number of harvester-mounted yield monitors remains relatively low and the ability to assess the scale of within-field variability using less invasive sensing technologies may prove useful on many farms.

The capability of remotely sensed crop reflectance information to support specific field-level analysis of within-season nitrogen recommendations in wheat and corn rotations (Flowers et. al., 2000), and the interpretation of management zones (e.g. Boydell and McBratney, 2002) is not new and is well documented. In addition, several derived vegetation indexes have been evaluated for assessing spatial variability of crop reflectance and crop yield predictions (e.g. Seelan et. al., 2003; Reyniers et. al., 2006; Haboudane et. al., 2007). Complimentary, pre-season information on variability in soil properties that often drive within-field crop variability have been mapped using apparent soil electrical conductivity (soil ECa) (Corwin et. al., 2003; Taylor et. al., 2003). Access to these data sources is widely available and their use in a quantitative index on production variability seems an obvious choice.

Methods

This work evaluates the suitability of remotely sensed imagery and proximal soil ECa sensors as a source of quantitative information for ranking the degree of within-field production variability found across farms. Parametric methods based on variogram analysis of the input data are used to quantify the degree of within-field variability as a function of the magnitude of variation (Mv) as well as the cohesion of spatial variability patterns relative to the present ability of variable-rate machinery to react (Sv) (Equation 1). A full description of the method for creating a yield-based opportunity index, Yieldex (Yi), can be found in de Oliveira et al. (2007).

Equation 1

This investigation applies the Yi methodology to a historical data set of crop reflectance imagery and soil ECa gathered on 7 farms from three grain grower groups in South Australia (SA), Victoria (VIC), and New South Wales (NSW). Fourteen broad-acre fields were mapped with crop yield monitors, from 1997 to 2006, and using electromagnetic induction (EMI) sensors and multispectral airborne imagery (AVNIR) from 2003 to 2006. Three depths of soil ECa observations (using EM31V, EM38V, and EM38H) and 10 vegetation indices (NDVI, GNDVI, PCD, PPR, PVR, VI, OSAVI, MSAVI, VI, TrVI) computed from the associated imagery were assessed. The best correlated vegetation index between imagery and interpolated yield data was chosen for each field for the process of determining the opportunity index from imagery. The opportunity indices computed from the imagery (Ii) and soil ECa (Si) were then compared with the Yi results calculated from the corresponding yield monitor data.

Results

The observed stable range of index values from the imagery and soil ECa as compared to that obtained from the crop yield data confirms the robustness of the process across data sources (Table 1). Table 2 shows a higher contribution from the magnitude component using the imagery data which is considered consistent with the finer resolution and response characteristics of the imagery data which provides more information on small scale variability. A higher observed contribution from the spatial structure component from the ECa data sets is consistent with the lower sampling resolution and the more continuous nature of soil properties being detected by the sensors.

Table 1. Yieldex (Yi) distributions from different data sources.

Index

Minimum

Median

Maximum

Yield (Yi)

1.6

5.5

20.2

Imagery (Ii)

2.1

6.5

18.4

ECa (Si)

2.2

4.7

9.0

Table 2. Yieldex (Yi) component correlations with the final index from different data sources.

Yi

Mv

Sv

Yield (Yi)

0.67

0.59

Imagery (Ii)

0.83

0.24

ECa (C)

0.47

0.83

Soil ECa Index (Si)

Correlations between the Si and the mean Yi calculated using the available years of yield data for each field did not show a strong relationship (r = 0.16) when all three sensors were incorporated (Table 3). However when analysed individually, the EM 38H shows up strongly as the best individual data source (r = 0.45).

Table 3. Correlations between mean yieldex (Yi) values for each field and apparent soil electrical conductivity (Si) values (measured by EM31V, EM38H, and EM38V) for all regions.

 

Field No.

Si (All EMI)
r

Si (31V)
r

Si (38H)
r

Si (38V)
r

Mean Yield Yi

14

0.16

0.03

0.45

0.09

A more detailed examination in each season across the three regions is shown in Table 4. It is clear that strong seasonality and regionality are present. The data from the EM38H (root zone soil ECa) provides the strongest correlations for the majority of years across the regions, while there appears an outstanding suitability for all the EMI sensors in the Riverine region.

Table 4. Correlations by year and region between Yieldex (Yi) values and apparent soil electrical conductivity (Si) (31V, 38H, and 38V) data

Year

SPAA

Riverine

CFI

(31V)
r

(38H)
r

(38V)
r

(31V)
r

(38H)
r

(38V)
r

(31V)
r

(38H)
r

(38V)
r

1999

-0.13

0.44

-0.40

-

-

-

-

-

-

2000

0.02

0.38

-0.27

0.25

0.19

0.53

-

-

-

2001

-0.32

0.37

-0.49

0.76

0.81

0.55

-

-

-

2002

-0.40

0.42

-0.51

0.60

0.62

0.78

-

-

-

2003

-0.06

-0.04

-0.20

0.93

0.91

0.79

0.04

0.08

-0.25

2004

0.60

0.15

0.49

0.82

0.87

0.78

-0.55

-0.79

-0.89

2005

0.03

0.74

-0.17

-

-

-

-0.07

0.58

0.31

Imagery Index (Ii)

Correlation of Ii with Yi for all field-year samples in the 3 regions shows a weak positive result (Table 5, r = 0.19) which is less than the results shown for the EM38H (r = 0.45). Across all field/years in each region, only the Riverine (r = 0.61) showed a positive result of the magnitude obtained from the Si. A further breakdown of correlations by year and by region shows results strongly varying among years and regions These range from strong negative correlations (2004 CFI, r = -0.77) to very strong positive correlations (2005 CFI, r = 0.99). These vast differences in responses appear to be related to seasonal moisture conditions, with 2005 having a relatively higher average rainfall across all regions. Obviously the Yi and Ii are seasonally dependent and the mid-season observation time of the imagery would mean that the impact of weather later in the season is unaccounted for. Adverse impacts on crop performance would be picked up in the Yi but not the Ii, potentially producing negative correlations between them. Individual vegetation indexes have shown correlation results from weak negative (r = -0.39 for MSAVI) to extremely high positive (r = 0.97 for VI) (data not shown) however these results are based on a very limited number of available field imagery data. A proper analysis by individual vegetation index will require more sampling years.

Table 5. Correlations between yieldex (Yi) values computed from yield data and the best vegetation index (Ii) by field.

Year

All Regions

CFI

Riverine

SPAA

Field No.

r

Field No.

r

Field No.

r

Field No.

r

2003

11

0.35

3

0

2

-

6

-0.44

2004

13

0.27

3

-0.77

4

0.49

6

-0.16

2005

7

0.34

3

0.99

-

-

4

0.49

2006

2

-

2

-

-

-

-

-

All Years

33

0.19

11

-0.08

6

0.61

16

-0.13

Discussion

Yieldex results are calculated from historical production data and so reflect the ‘opportunity’ captured in past seasons based on the associated crop type and management interactions. The relevance of Ii calculations to final yield harvested in a paddock appear to be very specific to paddock and season. The Si appears to be a more relevant indicator of the ‘opportunity’ realised in the final yield. Importantly, both new applications of the index have shown an ability to incorporate both the magnitude and spatial nature of encountered production variability in a manner that matches our understanding of the data produced by the respective sensing systems (Table 1). The Si and Ii show potential for use in situations where an assessment of the production variability between fields and farms is required, but where no yield monitor data is available. The Si may be more useful in a general assessment whereas the Ii would be useful in a single season comparison and particularly under conditions of good growing conditions. Such assessments would provide valuable information in ranking fields for further investment in site-specific management technologies. The use of these indices in crops other than broad acre would also be appropriate. The Si may be particularly useful in perennial cropping enterprises such as vineyards and orchards where the seasonal conditions and imagery capture requirements would cause less discrepancy with Yi.

Conclusion

The extraction of management information from fine-scale data monitoring activities is crucial to the adoption of PA. The accurate measurement of within-field variability and the ranking of the opportunity given by the quantity and patterns of variation would be useful to farmers contemplating further investment in site-specific crop management. Opportunity indices calculated from soil ECa and crop reflectance imagery have shown promise to support farmers in instances where spatially dense data on crop yield are unavailable.

References

Boydell B and McBratney AB (2002) Identifying within field management zones from cotton-yield estimates. Precision Agriculture, 3, pp. 9-23.

Corwin DL Kaffka SR Hopmans JW Mori Y van Groeningen JW van Kessel C Lesch SM and Oster JD (2003) Assessment and field-scale mapping of soil quality properties of a saline-sodic soil. Geoderma, V.114, pp. 231-259.

de Olivera RP Whelan BM McBratney AB and Taylor JA (2007). Yield variability as an index supporting management decisions: YIELDEX . Proceedings of the 6th ECPA, Skiathos, June 3-6, pp. 281-288.

Flowers M Weisz R and Heiniger R (2000) Aerial photographic determination of nitrogen application timing and rate recommendations in winter wheat. Proceedings of the 5th ICPA - International Conference for Precision Agriculture, Minneapolis, MN, July 16-19.

Fountas S Wulfsohn D Blackmore BS Jacobsen HL and Pedersen SM (2006) A model of decision-making and information flow for information-intensive agriculture. Agricultural Systems, 87, pp. 192-210.

Gangloff WJ Westfall DG Khosla R and Reich RM (2004). Areal association of production level management zone delineation and yield classes. Proceedings of the 7th ICPA – International Conference for Precision Agriculture, Minneapolis, MN, July 25-28.

Haboudane D Tremblay N Vigneault P and Miller JR (2007) Indices-based approach for chlorophyll content retrieval from hyperspectral data. Geoscience and Remote Sensing Symposium, IGARSS 2007, IEEE International, July 23-28.

McCown RL (2002) Locating agricultural decision support systems in the troubled past and socio-technical complexity of ‘models for management. Agricultural Systems, 74, pp. 11–25.

McBratney AB Whelan BM Ancev T and Bouma J (2005) Future directions of precision agriculture. Precision Agriculture, 6, pp. 1-17.

Pringle MJ McBratney AB Whelan BM and Taylor J. (2003).A preliminary approach to assessing the opportunity for site-specific crop management in a field, using a yield monitor. Agricultural Systems 76: 273-292

Reyniers M Vrindts E and de Baerdemaeker JD (2006) Comparison of an aerial-based system and an on the ground continuous measuring device to predict yield of winter wheat. European Journal of Agronomy, V.24, pp. 87-94.

Seelan SK Laguette S Casady MG and Seielstad A (2003) Remote sensing applications for precision agriculture: A learning community approach. Remote Sensing of Environment, V.88, pp.157-169

Taylor JC Wood GA Earl R and Godwin RJ (2003) "Soil Factors and their Influence on Within-field Crop Variability, Part II: Spatial Analysis and Determination of Management Zones." Biosystems Engineering, V. 84(4), pp. 441-453.

Whelan BM and McBratney AB (2000). The null hypothesis of precision agriculture management. Precision Agriculture 2: 265-279.

Previous PageTop Of PageNext Page

...

Search Site