Vous êtes sur la page 1sur 12

JAG

Volume

3 - Issue 2 - 2001

A wavelet characterization of high-resolution patterns for precision agriculture

NDVI

Virginie Epinatl, Alfred Stein 2, Steven M de Jong3 and Johan Bournal


1 Laboratory of Soil Science and Geology, Wageningen University, Wageningen, The Netherlands

2 Wageningen University, Mathematical and Statistical Methods group, Wageningen, The Netherlands and International Institute for Aerospace Surveys and Earth Sciences, PO Box 6, 7500 AA Enschede, The Netherlands (e-mail: alfred.stein@mat.dpw.wau.nl) 3 Wageningen University, Laboratory for geoinformation and remote sensing, Wageningen,

The Netherlands

KEYWORDS:

Wavelets,

NDVI,

Precision

Agriculture,

Spatial Statistics, Remote Sensing

ABSTRACT
This paper presents a quantitative analysis of patterns from airborne visible in

high-resolution ing. Attention procedure single field during

NDVI images obtained

remote

senspat-

focuses on the use of wavelets

to distinguish

terns of interest for precision agriculture for analyzing in the Netherlands,

at several scales. A general and applied to a different days resoof the images that is less releapproximation by at four

these images is presented monitored patterns decomposition

titular, multispectral high-resolution remote sensing may greatly contribute to precision farming by its possibility to assess leaf area development and crop cover at field scale during the growing season [Clevers, 19971. Using modern agricultural simulation models, one would then be able to translate these into management recommendations for the farmer, who may in turn apply an amount of nutrients and pesticides better suited to the conditions at the specific locations. Analysis of crop and soil condition patterns derived from remote sensing hence aims to improve decision making for precision agriculture. As current practices and modern technological equipment hardly allow for continuous changes in management, identification and quantification of soil and crop condition patterns is becoming of increasing importance. The use of remote sensing for precision agriculture requires a quantitative assessment of crop variables such as leaf area index (LAI) or cover [Moran et a/, 19971, characterization of the patterns, and in a later stage their integration over a period of time and their comparison. Patterns are here defined as the non-random distribution and arrangement of low and high values. In previous studies, fractals have been used to quantify vegetation patterns on remote sensing images [de Jong & Burrough, 19951. Also, geostatistical methods have successfully been used to quantify and compare spatial variability of properties of importance for precision agriculture [Bouma & Verhagen, 19981. In particular, Stein et a/ [I9971 used the cross-correlation function to compare spatial variability patterns of millet yield and soil data. Wavelets are a promising tool to quantify patterns by allowing pattern identification at various resolution levels and directions. Wavelets have been applied in the past to analyze images [Mallat, 19891. They have found many applications in remote sensing, such as removing speckle noise from radar images [Horgan, 19981, merging high spectral resolution images with high spatial resolution images [Yocky, 1996; Zhou et a/, 19981, and texture analysis and classification [Zhu & Yang, 19981. Here we use wavelets

one growing

season. Wavelet

is capable to reveal and quantify lution levels and directions vant for precision with different allowing agriculture

present at different Wavelet

and to filter information applications. approaches is useful within

wavelet functions

the backward-lookof the

ing and the forward-looking adaptation

of decision-making

of the analysis to the characteristics respectively.

available images or maps and to the possibilities of the existing sitespecific instruments,

INTRODUCTION In current practices of precision agriculture, management activities are guided by changes in crop and soil conditions that likely vary within the field. For instance, modern technological equipment allows changing application rates of agro-chemicals from place to place. In decisionmaking for precision agriculture, we commonly distinguish a backward-looking approach that aims at explaining differences which occurred during a growing season, from a forward-looking approach that aims at using the differences within the current growing season to support the farmer in decision making [Bouma et al, 19991. Within both approaches, an increasing need exists to quantify patterns of crop and soil conditions during the growing season. Remotely sensed images may be useful for this purpose, as geometric and spectral resolutions are being refined and the number of spectral bands is expected to further increase in the years to come [Goetz et al, 19851. In par-

121

Wavelet characterization

of high-resolution

NDVI patterns

JAG

Volume 3 - Issue 2 - 2001

to analyze agriculture different ticular of the observed turbing

remote as they

sensing allow

images

useful

for

precision into

May

30 and

July

11. Crop weight these

density

was

estimated

by meawith-

to decompose

the

image

counting the points.

plants

in a 1 m fraction of 10 plants

of several randomly

rows nearby selected

levels of resolution. level may be important pattern artifacts and to adjust whereas at other features,

The information to understand agricultural wavelets

at one parthe nature to disout equipment

Dry matter of 1 m from by correcting

and leaf area were locations.

sured for a sample in a vicinity biomass estimated per surface

From this, crop (LAI) were

may filter

unit and leaf area index by the plant density. point

levels. is to quantify derived a single a winter spatial patterns sensIn season. field

The objective

of this study scale during patterns during of

Final local yields and corresponding determined mounted made because the field. is REMOTE SENSING Remote Coupled Applications sensing Device observations imaging Airborne were on 3 with a grain mass flow harvester. dates: development on a combine different of the unequal

locations and 13

were a GPS were

of a set of four ing at the field particular, tures Netherlands sensitivity analyzed. patterns in the

NDVI images

from

remote

sensor 9,

growing to detect wheat wavelet season.

Measurements

we explore

use of wavelets

main feain the Moreover, function quantified

August

and

14

of the crop within

the 1997 growing we discuss how making

for the type of a particular Finally, in decision

to include

for precision

agriculture.

made with

the Dutch for

multispectral MATERIALS AND METHODS STUDY AREA Research situated west was conducted in Voorne Putten, on a 100 ha-commercial a former A single island winter wheat farm field of in the south-

airborne

scanner

CAESAR (ChargeScanner 19871 at an in the green

Experimental [Pouwels, was measured

in Remote

sensing)

altitude (550*15 (870*25 April rection in four 0.75 0.49 field

of 3 km. Reflectance nm), the red (670*15 nm) spectral was done reflectance

nm) and the near-infrared clear days in 1997: 7. Radiometric spectral corcharacterisresulted of between bands in the incident respectively, of were made in resolution spectral

of the Netherlands.

bands on four known with

14.7 ha (Figure of detailed viously nating from tent cent from yield

1) was selected data. This field deposits,

for the study based on its and availability preSoils, origias Typic varies con2 peron 9 has been described [1998]. classified Organic was are loam.

1, May 30, July 11 and August using the images m and

size, the variability

of soil characteristics & Stoorvogel

tics of panels laid on the ground. m x 0.75 m and

Pre-processing a spatial in the accuracy

by van Alphen marine

a geometric

Fluvaquents

[Soil Survey Staff, layer (25-30 and

19941. Their texture matter cm) varies between wheat between

1 .I1 m. Reflectance radiometer measures 2 m above

fine loam to heavy-clay in the upper and 10 percent.

670*10 11. This directed

nm and 870*10 instrument radiation band. vertically

nm was also measured simultaneously ground

using a hand-held

on May 30 and July using sensors

Winter

sown August

November and August terized Within

16, 1996, 14, 1997.

harvested

and reflected

The growing conditions. were

season was charac-

upwards

and downwards, measurements incident

by wet climatic the field,

for each spectral approximately located were covering (Figure measured vari1). on at 5 locations Calibration each spectral Plant cover Normalized as:

Its field of view

has a diameter

1 m. Replicate was done

nine points for detailed

in a radius of 2 m around by measuring both band with patterns Difference were sensors.

the nine points. radiation

ous soil conditions At these locations

measurements

crop characteristics

approached

by computing (NDVI)

the

Vegetation

Index

defined

NDVI = (NIR - R)/(NZR + R)


where tered

(1)

at 870

NIR is the reflectance in the spectral band cennm, and R the reflectance in the spectral
at 670 nm.

band centered WAVELETS We consider horizontal is denoted in both


2m cl 200 4mm

a digital and vertical by F&y),

image of a size M x N pixels in the direction with respectively. analysis, This image resolution a two-dimencombination of r the same spatial by a linear

directions. image

In a wavelet wavelet functions

sional

is approximated

FIGURE1: The study area and the sampling points.

two-dimensional

[Bruce & Gao, 19961.

122

Waveiet

characterization

of high-resolution

NDVI patterns

JAG

Volume

3 - Issue 2 - 2001

These wavelet functions are the simple product of onedimensional wavelet functions, one in the horizontal Xdirection and one in the vertical Y-direction. For each direction two one-dimensional wavelet functions exist: a wavelet function $(.) is used to represent smooth (low frequency) parts of signals and a wavelet function cp(.) their detailed (high frequency) parts. Consequently, four possible two-dimensional wavelet functions emerge,

ed in Figure 2 [Daubechies,

19921. The Haar wavelet

denoted by Q(x,y), cph(x,y>, rpv(x,y) and @(x,Y),


tively: @(XPY)= 4W44Y) @(su) = 4WrpM

respec-

(2)

P(X>Y) = cP(4.44Y) rpd(%Y) = rp(~).cpW where the superscript h denotes that rph(~,y) identifies features in the horizontal direction by capturing variation in the vertical direction by using r&). Similarly, rpv(~,y) identifies features in the vertical direction by using rp(x) and cpd(x,Y) in the diagonal direction. The function $(x,y) represents the smooth and low frequency part of the image. In this study, we used both the Haar wavelet, and the symmlet s8 constructed by Daubechies, present-

function 9(x) equals 1 for x E [O,l], and is 0 elsewhere, the function q(x) equals 1 for x E [0,1/z], -1 for x E [t/2,1], and 0 elsewhere. They have a compact support (zero outside a finite interval), are orthogonal and symmetric, but not continuous. Figures 2a-d show how the two-dimensional wavelet functions emerge from the one dimensional ones. Symmlet wavelets are defined in a similar way [Daubechies, 19921, are also orthogonal but have a compact support that is larger than that of the Haar wavelet, are continuous and nearly symmetric. Construction of the two-dimensional wavelet functions from the one-dimensional ones is illustrated in Figures Ze-h. Next, a first choice for the level of detail, indicated by the number of resolution levels J is to be made in advance. By definition, the resolution level of the original image equals 0, in this case corresponding to a spatial resolution r. Therefore, a resolution level j corresponds to a resolution of 2i.r. Then wavelet basis functions are defined with respect to a scaling factor 2-j forj = I,..., J and a translation vector (m,n) into the (x,y) direction. They are generated from the wavelet functions:

b(x)

4(Y)

&I

P(Y)

e FIGURE 2: Haar fky) (a),jhfx,yj (b), j$x,fi (c),j%,yJ (d) and symmlet fky) (e), jQ,y) (f), jv(x,y) (g), jd(x,y) (h) basis wavelet functions.

123

Wavelet characterization

of high-resolution

NDVI patterns

JAG

Volume 3 - Issue 2 - 2001

(3)

coarsest resolution level J. The smooth approximation of an image at a resolution level j-1 can be derived from the sum of the smooth and detail images at resolution level j:

s,_, (x, Y)
where

= S/(X?Y) + 107 (x7Y>


dir

di; is the direction (h, v or d). The wavelet approx-

imation F(x,y) of the image F(x,y) then equals the sum of two-dimensional wavelet functions at different scales and locations:

The smooth images S(x,y), j=l , .,J, are multi-resolution approximations of the image.
COMPARING WAVELET COEFFICIENTS The correlation coefficient smooth was used as a simple coeffiwas estiapproach to compare wavelet transform

where Sj,m,n and d,$& are the wavelet transform coefficients. Their absolute value can be considered as a measure of the contribution of the corresponding wavelet basis function to the image approximation. Coefficients Sj,m,n approximate the smooth part of the image at level j and difm,n, d&,and djfm,n represent at that level deviations from it into the different directions. Wavelet functions are applied at a broad range of scales (from 1 to J) at a large number of pixels (for each combination of m and n) and into 3 directions. Notice that coefficients s only occur at level J and that smooth variation at lower levels j is captured by coefficients d at level j+Z. A wavelet approximation includes the smooth function only at level Jand the detailed wavelet functions at all levels. The coefficients Sj,m,n are given by: (5) A similar expression applies to the coefficients d$&. In the practice of image analysis, integrals are replaced by finite summations. The coefficients are grouped into socalled crystals, also referred to as subbands [Daubechies, 19921, according to resolution level j and, for coefficients d, direction. Contribution of the separate crystals to the reconstructed image is measured by the percentage of energy by crystal, defined as the sum of squares of the coefficients of that crystal, divided by the sum of squares of the pixel values of the whole image. Anisotropic autocorrelation of the coefficients into different directions reveals the presence of spatial structure at various scales. The two-dimensional wavelet approximation (Equation 4) can also be expressed as the sum of 3*J+l image components corresponding to different resolution levels and different directions:

cients at a given resolution mated

level. The coefficient

shifted over the two crystals, where one pixel represents a wavelet coefficient at position (m,n), and assigned to the central pixel.

using a 5x 5 pixel window

RESULTS
NDVI IMAGES Figure 3 presents four rotated NDVI images derived from the CAESAR observations and Table 1 gives their descriptive statistics. Figure 4 represents the development stage of the crop at the image acquisition dates. Values of 0, 1 and 2 indicate emergence, flowering and maturity of the crop, respectively. The first two images were taken during the vegetative phase while the last two images correspond to the reproductive phase. Between the second and the third observations, a disease caused by a fungus (fusarium) broke out and lodging occurred. As can be seen in Figure 3, all images show management features, like tracks in the 21 direction with the EW direction, old field boundaries and a net of former canals perpendicular to these. On the image of April 1 (NDVI=0.33*0.074), three main regions can be distinguished. Region A located at the upper part of the field presents higher NDVI values, region B at the lower part shows lower values, and region C at the lower left part of the field has again higher values (Figure 1). The boundary between regions A and B corresponds to the border of former parcels with different land uses: grassland on A and cropland on B. On May 30 and July 11 NDVI values are very high (0.983kO.026) and (0.920*0.020), respectively, indicating that the soil is in most places fully covered by the crop. Regions A and B can still be distinguished, but region C is now absent. The image on July 11 shows a pattern opposite to the pattern on April 1, with slightly different regions. Region A presents lower values especially at the left of the field, region B higher values and region C shows again lower values. The lower values in region A can be explained by the breakout of fusarium and lodging in that part of the field. Two parallel black lines correspond to missing spectral values. On August 7, just before harvesting, NDVI values have reduced to 0.459kO.074, indicating

(6)
where Sj(X,y) is the smooth image and ~,?(x,y) are detail images at resolution level j, showing object edges in the horizontal, the vertical and the diagonal direction. Sj(x,y) is a smooth approximation of the image at the

124

Wavelet

characterization

of high-resolution

NDVI patterns

JAG

Volume

3 - Issue 2 - 2001

~~~

0
a

2;0

, O 200 400 600

I800

400

600

800

090

092

094

096

098

1
100

, 0
c

200

400

600

Em

200

400

600

800

FIGURE 3:

NDVI images on April 1 (a), May 30 (b), July 11 (c), August 7(d).

TABLE 1:

Summary statistics of Caesar NDVI on April 1, May 30, July 11, August 7 (n = 261238)
April Minimum 1st quartile Median Mean 3rd quartile Maximum Standard deviation 0.029 0.269 0.323 0.327 0.373 0.763 0.074 7 May 30 0.480 0.979 0.989 0.983 0.996 1 .ooo 0.026 July 11 0.553 0.912 0.925 0.920 0.933 0.999 0.020 August 0.252 0.404 0.460 0.459 0.510 1 .ooo 0.074 7

I-----l

I)-

that discoloring of the crop has started. The three same zones can be distinguished with higher values at the upper part, lower at the lower part, and lowest at the lower left part of the field.
COMPARISON SUREMENTS BETWEEN AND HAND-HELD RADIOMETER MEA-

FIGURE 4: Development stage of the crop at the image acquisition dates (represented by arrows)

CROP CHARACTERISTICS

Summary statistics of the crop characteristics presented in Table 2 show that the average plant density almost did not vary between May 30 and July 11. Above-ground biomass increased during this period, mainly due to the development of the ears. However, the average green LAI decreased, due to the discoloring of the crop that is

on July 11 in its ripening phase (Figure 4). Table 3 shows that NDVI derived from hand-held radiometer measurements is significantly correlated with plant density at the 0.05 level on May 30, and at the 0.1 level on July 11. No clear relation between this vegetation index and the LAI or above-ground biomass can be found, probably because a complete soil cover by the crop is reached in most places, and because of the high LAI values (Table 2). As a result, the NDVI curve tends then to saturate and

125

Wavelet

characterization

of high-resolution

NDVI patterns

JAG

Volume 3 - Issue 2 - 2001

TABLE 2: Summary statistics of crop characteristics held radiometer on May 30 and July 11 (n = 9) Minimum NDVI Plant density (plantslmz) May 30 Green LAI Above-ground biomass (g/mz) NDVI Plant density (plants/mz) July 11 Green LAI 0.723 152 3.72 559 0.666 144 Mean 0.743 186 6.25 785 0.679 183 Maximum

and handStandard deviation 0.011 20 1.92 155 0.009 25

0.757 213 9.54 1079 0.697 212

2.75 1298

4.14 1771

6.34 2261

1.07 259

Above-ground biomass (g/m2)

/
0

/
200

400

600

800

FIGURE5: Yield map (t/ha)

TABLE 3: Correlation between NDVI derived from handheld radiometer and crop characteristics on May 30 and July 11
May 30 Plant density Green LAI Above-ground
*significant

the

area where

the

disease

and

lodging

occurred,

and

higher probably physical

yields in the southern errors due to flow might of the to and to

part of the field. processes within

Local lower are most the yield comto mearelated

values at the beginning bine. Other surement moisture changes outliers system, content, and varying
ANALYSIS

and end of the transects be caused combine the by errors and the such such

July 11 0.646* 0.533 0.241


level.

0.690** 0.106 biomass -0.204


significant at the 0.05

properties

in-field cutting

differences operator [Thylen width

as crop as speed

at the 0.1 level

; **

et al, 19971.

NDVI

variations between

within

the

field

tend

to

be

small. on the ie,

WAVELET

Discrepancy side and the other

in correlations side can

between

NDVI and LAI on one biomass fact that material, by the

For the wavelet angle ment direction, tracks. chosen,

analysis that

we rotated horizontally direction

the images

over an but levof 48 coeffias

NDVI and above-ground be explained includes plant non-green

of 21, thus setting

the main manageof dominant were assigned

is the

above-ground throughout Summary

biomass the vertical

irrelevant els were

Pixels outside to avoid which

the field

dead leaves, stems and ears, scattered profile.

over several

layers

the average

NDVI value

bias. Six resolution resolution transform wavelet

give a coarsest approximations

m on the ground. statistics of the NDVI derived smaller from from the handheld radiometer deviations which tions metric the between than observations those derived the at the 9 points on May 30 ed with

In this study, wavelet

cients and multi-resolution the symmlet well as with

were comput-

s8 for all NDVI images, reconstruction by reflecting of the image transform image with

and July 11 (Table 2) show Possible explanations reflectance measured,

means and standard the CAESAR images. number of points at radiometer was

the Haar wavelet were reduced

for the image of July 11, In and artifacts the original them. by at was done coefficients the

the first case, errors in image at its boundaries In the second assigning image the boundaries of July observable
July 11

are the limited using

hand-held

series at the boundaries

and periodically

extending

the differences

in bandwidth resolution

of the observain radioto relations according that

case, extension of the image. 11, being the

in the red and near-infrared corrections sensing hand-held and in spatial device. radiometer remote

wavelengths,

zeros to the wavelet

We now first focus on the clearest

We assumed

NDVI and crop character-

pattern.

istics are also valid for CAESAR NDVI images.


YIELD MAP

We notice a rotated point data yield with map derived the nearest from the neighbor a. Horizontal

for the wavelet with mainly

transform

coefficients that:

result-

Figure original method with 1.93 their

5 presents 12,449 applied

ing of a decomposition crystals levels level

the s8 wavelet depict

to a 0.75 m resolution data set vary between of the

grid. Yield values deviation points of to

the tracks, in region regions 6).

especially as the become A. When

in the original t/ha.

0.1 t/ha and 22 t/ha,

at resolution area with moving more from

between 4 to while level

2 and 4, as well 5 the (Figure

an average nearest

of 8.48 t/ha and a standard measurement 2.11 is on average

the disease

and lodging

The distance neighbor

m and varies yields at

pronounced,

tracks disappear,

as the scale

between

0 m and 3.92 m. The map shows lower

of resolution

becomes

coarser

126

Wavelet

characterization

of high-resolution

NDVI patterns

JAG

Volume

3 - Issue 2 - 2001

b.Vertical crystals mainly reproduce the disease zone in region A at all resolution levels, as well as the two lines at levels 1 to 3 and the boundary of older parcels at the finest levels. c. Diagonal crystals capture in general less variation than horizontal and vertical crystals. Higher coefficients are located at region A. Old boundaries can be distinguished at resolution level 2. Pattern features that are of minor importance for precision farming are therefore most prominent at the finest levels, and mainly into horizontal and vertical directions. The interesting change in resolution occurs between levels 3 and 4 where the pattern of the main regions is now more prominently being quantified. We will now focus on levels 3 and 4. Wavelet transform coefficients and autocorrelations are plotted in Figure 7 for level 3 and in Figure 8 for level 4, corresponding to 6 m and 12 m on the ground, respectively. At level 3 we notice that horizontal crystals present clear spatial correlation in the horizontal direction, and have a periodicity of approximately 75 m in the vertical direction, mainly due to presence of tracks. The vertical crystal shows a pattern with strong correlations in the vertical direction, and alternating correlations in the horizontal direction. This is probably caused by the two lines and the field boundary. The remainder of the pattern is composed of higher coefficients in region A with variation that is mainly noise, as represented by an almost flat correlation function. The diagonal crystal shows mainly noise, with some pattern visible in region

A, leading to an almost flat autocorrelation

function.

At level 4, the vertical crystal now shows some more structure, with clear dark areas at the top left, bottom left and right parts of the crystal (Figure 8). This more prominent pattern type is present as well in an autocorrelation function that shows more structure than at level 3. On the other hand, the diagonal crystal shows some structured variation, that, however, is not represented by the autocorrelation function, that mainly remains flat, but alternates in the horizontal direction for small distances. We next turn towards the multi-resolution approximations with the s8 wavelet (Figure 9). At the coarsest resolution level (level 6), the 3 main zones corresponding to regions A, B and C occur. From resolution level 5 (- 24 m) onwards, a grid-like structure appears. This structure refines and is most obvious at resolution level 4 (- 12 m). Differences in image texture between regions A and C occur from level 3 onwards. We have seen that the wavelet analysis using the s8 wavelet functions reveals patterns such as the regions A, B and C that are of primary interest for precision agriculture. These regions might display differences in crop conditions that can be treated in a location specific way. Moreover, patterns that are visible but are caused by objects that are less relevant to precision farming are filtered out from a specific level of detail onwards. Use of the s8 wavelet, therefore, contributes to analyzing processes that are relevant for precision agriculture.

FIGURE 6: Matrix of absolute s8 wavelet transform coefficients of the NDVI image of July 11 in horizontal direction at resolution levels 1 (a), 2 (b), 3 (c), 4 (d), 5 (e) and 6 (6.

127

Wavelet characterization

of high-resolution

NDVI patterns

JAG

Volume 3 - Issue 2 - 2001

Sensitivity to the type of wavelet Wavelet that analysis the was also carried Haar wavelet distribution contribute out (Figure for the image of July 11 using similar wavelet, 10). We found as for the s8 into crystals dif-

(Table 4). Multi-resolution Haar wavelets easily allow (Figure the present for recognition error

approximations a blocky artifacts pattern

obtained that

with

does not as A, B and C with and is image 0 though

of the 3 regions the original level

patterns

for all crystals emerge of energy type, more

IO), and fewer

than approximations

although and

~8. The relative reconstructed now lower

L2 between

fers according

to wavelet

as coefficients to the total

dtM,n,
energy

image

at resolution

da,,,

dim,,

(2 x 1 O-15) as compared

to 3 x 1O-3 for ~8.

e (_
, ,

;
,

Vi ,,

:
.i -; . ,. :

_ : .

,I Iii,
,I ,,

.
,_
l:__.

:;,
\..
_:.

,
j
!

n
b f

FIGURE7: Matrix of absolute s8 wavelet transform coefficients and corresponding autocorrelation July 11 at resolution level 3 (12 m) in horizontal (a, b), vertical (c, d) and diagonal (e, f) directions.

functions of the NDVI image of

d functions of the NDVI image of

FIGURE8: Matrix of absolute s8 wavelet transform coefficients and corresponding autocorrelation July 11 at resolution level 4 (12 m) in horizontal (a, b), vertical (c, d) and diagonal (e, f) directions.

128

Wavelet characterization of high-resolution NDVI patterns

JAG l Volume 3 - issue 2 - 2001

Comparison

between

yield

and

My

11 NDVi

smooth

wave/et coefficients

Yield and NDVI smooth wavelet transform coefficients were compared at identical resolution levels by estimating the correlation coefficient within a 5 x 5 pixels-window shifted on the crystals. Figure 11 presents the correlation coefficient for comparison of the yield map and

the original NDVl image at level 3. Clear positive correlations exist in the upper part of the image, whereas strong negative correlations exist in the southern part of the image, indicating the similarity of the overall yield and NDVI patterns in the first location, and the inversion in the second location.

FIGURE9: Approximations

of the NDVI image of July 11 with s8 at resolution levels 6 (a), 5 (b), 4 (c), 3 (d), 2 (e) and 1 (f).

FIGURElo: Approximations of the NDVI image of July 11 with Haar at resolution levels 6 (a), 5 (b), 4 (c), 3 (d), 2 (e) and 1 (f).

129

Wavelet

characterization

of high-resolution

NDVI patterns

JAG

Volume

3 - Issue 2 - 2001

TABLE 4: Percentage of energy by crystal for NDVI images on April 1, May 30, July 11 and August 7 April 1 s8 d: dl" df d! ds" d2" d3" di' d: d4" di' d4" dsh ds" d: dsh d8' dff
S6

reflecting ments that show

natural

soil

variability resolution for resolution approach appears suited

or

historical agriculture these

developclearly in

are important coarser allowed

for precision levels. detecting

May30 58 0.0023 0.0015 0.0011 0.0042 0.0019 0.0003 0.0049 0.0035 0.0004 0.0054 0.0044 0.0004 0.0030 0.0011 0.0004 0.0014 0.0012 0.0001 99.9624

July 11 s8 0.0005 0.0006 0.0002 0.0019 0.0013 0.0001 0.0041 0.0019 0.0004 0.0040 0.0015 0.0005 0.0026 0.0009 0.0003 0.0009 0.0003 0.0002 99.9777

August 7 s8 0.0377 0.0183 0.0149 0.1029 0.0167 0.0041 0.1264 0.0311 0.0077 0.1087 0.0458 0.0106 0.0974 0.0487 0.0106 0.0998 0.0343 0.0126 99.1718

July 11 Haar 0.0014 0.0010 0.0003 0.0030 0.0016 0.0003 0.0046 0.0014 0.0004 0.0036 0.0017 0.0005 1.0395 0.7685 0.0074 0.4876 0.0008 0.0003 97.6759

up at

Multi-resolution features

approximations the images

0.0298 0.0163 0.0103 0.0844 0.0242 0.0055 0.1824 0.0425 0.0129 0.1556 0.0717 0.0161 0.1714 0.0673 0.0174 0.1662 0.0475 0.0243 98.8543

at different

levels. for decision making, as its of conapproxiat difor

For the backward-looking the symmlet wavelet makes smoothness tinuous mations ferent that it well

to be most

useful,

for the analysis comparisons in plant

patterns allow resolution

such as NDVI. Multi-resolution to make quantitative levels. Differences by differences and

development therevaricaused seaa maps of

could

be caused are observed scale specific

in soil conditions Wavelets different have

elevation to relate better with

quantified. between

fore provide

a tool to quantify patterns of

these patterns,

and hence

ables or different these patterns. similar son or between combination derived In the wavelets from that

days. As such their processes during

use may lead to a that of areas in the field the growing benefit and yield the property

understanding plant

Also, characterization development growing sensing harvester. approach, seasons

might images

from

of remote a combine they

forward-looking

approximate

original resolutions

remote

sensing does not

images at a range of different


Other images

can be used to multi-resotranslate comparing on remote and as such to the actual technology.

the advantage. 7, s8 wavelet some tracks level 2 visible transform by the horionwards. on vertiallow lution On the emerging sensing mation (not shown) from depict

Precision

agriculture using the

technology yet. Hence, Haar wavelet allow

On April coefficients zontal Boundaries

1, May 30 and August crystals resolution

for continuous approximations other plant images. hand,

applications

NDVI images to the precision the growth Observed

of the available as observed on the field practices

of older

parcels are especially crystals contain, crystals. lower A) from 11, the

coefficients patterns

cal crystals of the first and the last images, levels 3 to 5. Diagonal high coefficients of the (regions upper image of July than the other

at resolution

patterns pattern

as on July 1 1, less On the opposite part of the field level 2 auto-corpresent a in than the at resolu-

are a coarse approxi-

to the cropping

may help to adapt conditions.

management

B and C) present

more high coefficients resolution crystals horizontal crystals in horizontal a periodic in vertical Auto-correlation

part of the field (region

on May 30. Coefficients tion levels 3 and 4 present relation, vertical while coefficients crystals. auto-correlation.

is negligible

the diagonal Approximations m and between and cdals tion&

at resolution ground) A and region

levels 6 and especially show a clear B on the images due to tracks, on all images

5 (48
a

24 m on the region

separation of April at a resolu1 parcels

and August

7. The pattern

older

Y
Y

is the most obvious

12 m.
Y .
1

DISCUSSION A wavelet be a good lution tures were

AND

CONCLUSIONS of the four to quantify at fine NDVI images proved to


E

transform method

patterns Historical resolution

at different management levels.

resofea-

FIGURE II:

Smooth

wavelet

transform

levels and directions. separated

Patterns

and July 11 NDVI (b) at resolution correlation coefficient (c).

coefficients of yield (a) level 3 and corresponding

130

Wavelet

characterization

of high-resolution

NDVI patterns

JAG

Volume

3 - Issue 2 - 2001

In comparison with other pattern analysis techniques,

the

main benefits of a wavelet approximation are that several resolutions and directions are split out, and help to characterize an individual image. Fractals quantify variation over a range of scales but do usually not allow for discrimination between variations occurring at different scales. Geostatistics are useful for interpolation of individual variables and making univariate and bivariate spatial dependence functions. Moreover, geostatistical simulations are useful in analysis of error propagation. Though, a decomposition of an image as in this study is outside the range of current geostatistical methods. Wavelets being a modern tool for spatial analysis provide in a way a useful addition to geostatistical methods. Using the wavelet characterization of high-resolution NDVI patterns allows to make a quantitative comparison at different scales between images and hence between variables that reveal variation at different scales. If, for example, an NDVI image mainly shows variation at one particular scale, corresponding to a specific resolution j, then quantifying the relation with other variables and images showing variation at the same scale may lead to useful insight into causes of spatial variability. Such a wavelet characterization therefore quantitatively relates variability between different images, instead of entirely focusing on average values that have been used so far. This is a step forward in precision agriculture where the main focus is within-field variability.

Daubechies, I., 1992. Ten Lectures Industrial and Applied Mathematics,

on Wavelets. Society for Philadelphia, PA, 357 pp.

de Jong, S.M. & P.A Burrough, 1995. A fractal approach to the classification of mediterranean vegetation types in remotely sensed images. Photogrammetric Engineering and Remote Sensing 61: 1041-1053. Goetz, A.F.H., G. Vane, J.E. Solomon & B.N. Rock, 1985. Imaging spectrometry for earth remote sensing. Science 228: 11471153. Horgan, G., 1998. Wavelets Photogrammetric Engineering 1177. for SAR image smoothing. and Remote Sensing 64: 1171-

Mallat, S.G., 1989. A theory for multiresolution signal decomposition: the wavelet representation. IEEE Transactions on Pattern Analysis and Machine intelligence 1 1: 674-693. Moran, M.S., Y. lnoue & E.M. Barnes, 1997. Opportunities and limitations for image-based remote sensing in precision crop management. Remote Sensing of Environment 61: 3 19-346. Pouwels, H., 1987. User guide to CAESAR. BCRS, Delft. Soil Survey Staff, 1994. Keys to Soil Taxonomy, NRCS, Washington, DC, 306 pp. 7th Edition. USDA-

Stein, A., J. Brouwer & J. Bouma, 1997. Methods for comparing spatial variability patterns of millet yield and soil data. Soil Science Society of America Journal 61: 861-870. Thylen, L., P. Jijrschik & D.P.L. Murphy, 1997. Improving the quality of yield data. In: J.V. Stafford (Ed.). Precision Agriculture 97: papers presented at the first European conference on precision agriculture, Warwick University Conference Center, UK, 7-l 0 September 1997. BIOS, Oxford, pp. 743-750. van Alphen, B.J. & J.J. Stoorvogel, 1998. A methodology to define management units in support of an integrated, model-based approach to precision agriculture. In Robert, P.C., Rust, R.H., and Larson, W.E. (Eds.). Proceedings, Fourth International Conference on Precision Agriculture, 19-22 July 1998, St. Paul, Minnesota. ASA-CSSA-SSSA, Madison, Wisconson, pp. 12671278. Yocky, D.A., 1996. Multiresolution wavelet decomposition image merger of Landsat Thematic Mapper and SPOT panchromatic data. Photogrammetric Engineering and Remote Sensing 62: 1067-I 074. Zhou, J., D.L. Civco & J.A. Silander, 1998. A wavelet transform method to merge Landsat TM and SPOT panchromatic data. International Journal of Remote Sensing 19: 743-757. Zhu, C. & X. Yang, 1998. Study of remote sensing image texture analysis and classification using wavelet. International Journal of Remote Sensing 19: 3197-3203.

ACKNOWLEDGEMENTS This study is embedded within a large-scale research project funded by the Dutch Board for Remote Sensing (BCRS). The research of V. Epinat is funded by the Earth and Life Sciences Research Council of the Netherlands Organization for Scientific Research (NWO-ALW). The van Bergeijk family members are kindly acknowledged for supporting this research on their farm.

REFERENCES
Bouma, J. & J. Verhagen, 1998. Environmental threshold values for agricultural production systems varying in space and time. In: V. Barnett, A. Stein & K.F. Turkman (Eds.), Statistics for the Environment 4: Statistical Aspects of Health and the Environment. John Wiley, Chichester, pp. 293-303. Bouma, J., J. Stoorvogel, B.J. van Alphen & H.W.G. Booltink, 1999. Pedology, precision agriculture and the changing paradigm of agricultural research. Soil Science Society of America Journal 63: 1763-l 768. Bruce, A. & H.Y. Gao, 1996. Applied wavelet Springer Verlag, New York, NY, 338 pp. analysis with Splus.

RESUME
Cet article presente une analyse quantitative de formes visibles

dans des images NDVI obtenues


aerienne. Lattention lettes pour distinguer

a partir de la teledetection est concentree sur Iutilisation dondedes formes presentant un inter& pour

Clevers, J.G.P.W., 1997. A simplified approach for yield prediction of sugar beet based on optical remote sensing data. Remote Sensing of Environment 61: 221-228.

une agriculture de precision a plusieurs echelles. Une procedure g&Wale pour analyser ces images est presentee et appliquee a un seul champ aux Pays-Bas, controlee durant quatre jours pendant une saison de croissance. La decomposition par ondelettes de ces images est capable de reveler et quantifier des formes presentes a differents niveaux de resolution et directions et de filtrer Iinformation qui est moins importante pour des applications dune agriculture de precision Une approximation par ondelettes avec differentes fonctions dondelettes est utile dans les approches de decision avec un regard vers ie passe et un regard vers Iavenir en permettant une adaptation de Iana-

131

Wavelet

characterization

of high-resolution

NDVI patterns

JAG

Volume

- Issue 2 - 2001

lyse aux caracteristiques des images ou cartes disponibles et aux possibilites des instruments existants specifiques du site, respectivement.

RESUMEN
Este articulo presenta un analisis cuantitativo de 10s patrones visibles en imdgenes NDVI de alta resolution obtenidas por teledetection aeroportada. La atencion se concentra en el use de pequenas ondas (wavelets), para distinguir patrones de inter& en agricultura de precision a diferentes escalas. Se presenta un procedimiento general para analizar estas imagenes; el mismo ha sido aplicado a una parcela agricola en 10s Paises Bajos, la cual se monitoreo en cuatro diferentes dias durante una esta-

cion de crecimiento de 10s cultivos. La descomposicion de las pequenas ondas en las imdgenes es capaz de revelar y cuantificar patrones presentes a diferentes niveles de resolution y en diferentes direcciones, y de filtrar la information que es menos relevante, para aplicaciones en agricultura de precision. Esta aproximacion basada en diferentes funciones de ondas pequeAas es titil en 10s enfoques regresivos y progresivos de la toma de decisiones, porque permite adaptar el andlisis a las caracteristicas de las imdgenes o de 10s mapas disponibles y a las posibilidades de 10s instrumentos existentes en 10s sitios especificos, respectivamente.

132

Vous aimerez peut-être aussi