Combined use of optical and microwave remote sensing data for crop growth monitoring

ELSEVIER

Combined Use of Optical and Microwave Remote Sensing Data for Crop Growth Monitoring j. G. P. w. Clevers* and H. J. C. van Leeuwen* I n this study, optical and microwave remote sensing data were used in combination for crop growth monitoring. A simple reflectance model was used for estimating leaf area index (LAI) from optical data, and a simple backscatter model was used for estimating LAI from radar data. Subsequently, the synergistic effect of using both optical and radar data for estimating LAI was analyzed by studying different data acquisition scenarios. Finally, the remote sensing models were inverted to obtain LAI estimates during the growing season for use in calibrating the crop growth model to actual growing conditions. This concept for crop growth monitoring is elucidated and illustrated with examples using ground-based and airborne data obtained during the MAC Europe 1991 campaign. Results showed that simultaneous optical and radar observations did not improve estimates of LAI over optical data alone. However, for operational applications the assumption of nonsimultaneous observations is more realistic. Results for sugar beet indicated that when periodic (about every ten days) optical recordings are available throughout most of the growing season, LAI can be monitored well and a good estimate of sugar beet yield at the end of the season is possible by using a calibrated crop growth model. When only a few recording dates with an optical sensor are available, radar recordings at L-band HH polarization or C-band VV polarization gave a slight improvement of the results of crop monitoring and yield estimation compared with the optical data alone. In the absence of optical remote sensing data, radar data yielded a significant improvement in yield

Department of Landsurveying and Remote Sensing, Wageningen Agricultural University, Wageningen, The Netherlands Address correspondence to J. G. P. W. Clevers, Dept. of Landsurveying & Remote Sensing, Wageningen Agricultural Univ., P. O. Box 339, 6700 AH Wageningen, The Netherlands. Received 10 May 1995; revised 28 September 1995. REMOTE SENS. ENVIRON. 56:42-51 (1996) ©Elsevier Science Inc., 1996 655 Avenue of the Americas, New York, NY 20010

estimation compared with the case of no remotely observed information. This confirmed that the main advantage of radar lies in acquiring information on crop growth when other techniques (in particular optical techniques) fail.

INTRODUCTION In agricultural planning and policy making, for example, in the European Union, knowledge of crop production at an early stage is very important at both national and regional levels. The two constituents of crop production are crop acreage and crop yield. Better estimates of crop yield are obtained if the growth of the crops is being monitored during the growing season. The crop variable leaf area index (LAI) is important as a measure for crop growth. Crop growth can be monitored by using crop growth models. However, estimates of crop growth often are inaccurate for nonoptimal growing conditions. Remote sensing can provide information on the actual status of agricultural crops, thus calibrating the growth model for actual growing conditions (Maas, 1988; Bouman, 1991; Del~colle et al., 1992). Best results are obtained by using (reflective) optical remote sensing data (e.g., a vegetation index) in estimating the LAI regularly during the growing season and subsequently calibrating the growth model based on periodic LAI estimates (Clevers et al., 1994). However, at national and regional scale in, for example, Europe, the regular acquisition of optical remote sensing data is hampered by frequent cloud cover. Radar can acquire remote sensing information with a high temporal resolution due to its all-weather capability. Moreover, data from the optical and microwave windows provide complementary information and the combined use, either contemporaneously or at dif0034-4257 / 96 / $15.00 SSDI 0034-4257(95)00227-8

Combined Use of Optical and Microwave Data

43

8-

ferent times during the growing season, can improve the estimation of crop variables. It is studied whether yield estimates from both windows together (called synergy) are more accurate than estimates from each window separately. In this article, LAI is the essential link between optical and radar remote sensing techniques and crop growth models. The LAI is estimated with the derived inverse remote sensing models and brought in the calibration process of the crop growth model with the appropriate weighting factor. This concept is illustrated with data from the MAC Europe 1991 campaign over the Dutch test site Flevoland.

5~4

DIRECT AND INVERSE MODELING IN REMOTE SENSING

Figure 1. Relation between estimated LAI using the CLAIR

LAI Estimation with the CLAIR Model A simplified, semiempirical reflectance model for estimating LAI of a green canopy was introduced by Clevers (1988; 1989). It is called the CLAIR model. In this model, first, the WDVI (weighted difference vegetation index) is ascertained as a weighted difference between the measured NIR and red reflectances, assuming that the ratio of NIR and red reflectances of bare soil is constant. In this way a correction for the influence of soil background is performed:

WDVI = NIR - C" Red,

(1)

where NIR = measured NIR reflectance, Red = measured red reflectance, C = slope of the (soil-specific) soil line, or ratio between NIR and red reflectance of soil. Subsequently, this WDVI is used for estimating LAI according to the inverse of an exponential function: LAI = - 1 / a'ln(1 - WDVI / W D V L )

(2)

with a as a parameter describing the rate with which the function of Eq. (2) runs to its asymptotic value and W D V L is the asymptotic limiting value for the WDVI. The exponential relationship between WDVI and LAI means that LAI estimates will be less accurate when approximating the asymptotic value of WDVI (WDVL). In other words, the accuracy of LAI estimation will decrease with increasing LAI value. A firstorder approximation of the standard deviation of LAI estimation can be derived as tr(LAI) = exp(a. LAI - ln(a" WDVL)) •a(WDVI).

(3)

Bouman et al. (1992) validated the CLAIR model for sugar beet, and they estimated the value of a to be 0.485 and the value of W D V L to be 56.3. The residual mean square for the calibration set was 4.1 (in terms of squared WDVI units). This value may be used as an

o o o o o AGRISCATI data __ LAI + 2*sigma __ LAI - 2*sigma sigma (WDVI)=2

_o6

/ /

121

E

-'o"°

%20

/.

/

4J

o

,1/- 7 7

o

2

4 LAI m e a s u r e d

6

model and measured LAI for sugar beet. Flevoland test site, AGRISCATT campaigns 1987 and 1988.

estimate of the variance of the individual WDVI measurements. The resulting estimate of a(WDVI) in Eq. (3) is 2.0. Figure 1 plots the estimated LAI using the CLAIR model against the measured LAI (ground measurements) for the calibration set used by Bouman et al. (1992). In addition, the lines defining 2 standard deviations from the measured LAI are shown. LAI Estimation with the Cloud Model In former studies (Hoekman et al., 1982; van Leeuwen, 1992) it was shown that the Cloud model (Attema and Ulaby, 1978) is a simplified, semiempirical radar model, which is useful for explaining the radar backscatter of agricultural crops. However, in general, this model is only valid during the beginning of the growing season, because after closure of the crop a constant backscatter level is reached. Another limitation is the calibration and validation of the model. A high temporal resolution is needed for calibrating the radar model within this short period of time (e.g., sugar beets take 3-4 weeks from bare soil to closure). The original formulation of the Cloud model assumed that the vegetation consists of a collection of water droplets, which are represented as small identical particles. Equation (4) in the original formulation by Attema and Ulaby (1978) is

y = C" (1 - exp( - D' W" h / cos 0)) + G" exp((B'm, - D" W . h) / cos O).

(4)

The first term represents the vegetation contribution to the measured scattering cross section 7 (the radar cross section per unit of area projected in the direction of transmission), and the second term represents the soil contribution. The parameter C represents radar backscatter at full closure of the canopy, G is the dry soil characteristic with roughness information incorporated,

44

Clevers and van Leeuwen

D is the two-way attenuation of the radar wave through the canopy, and B is the sensitivity of backscatter to soil moisture. W represents the water content of the canopy, h equals the canopy height, and ms is the volumetric moisture content of the soil. Finally, 0 is the incidence angle of the radar beam. For sugar beet a constant relationship (factor A) between the amount of crop moisture (W.h) and the LAI was found: LAI : a" W. h,

and

L-Band HH 0.8967 0.1369 0.1767 0.6250 0.000505 0.022

D' parameter C parameter K parameter R2 Residual mean s q u a r e o(T)

C-BandVV 0.3660 0.6821 0.4394 0.6665 0.003012 0.055

(5)

where A had the value 0.83 m 2 / kg (van Leeuwen, 1996). For one date in the growing season we may consider the soil moisture content (ms) and the soil roughness for all sugar beet fields in Flevoland constant. If we put K=C-G'exp(B'ms)

Table 1. Calibration Results of the Cloud Model in Eq. (6) for Sugar Beet Using Data from MAC Europe 1991 on the Flevoland Test Site

D'=D/A,

the Cloud model can be inverted and rewritten as LAI= - c o s O / O " l n ( ( y - C ) / -K).

C o m b i n e d Use of C o n t e m p o r a n e o u s Optical and Radar D a t a When looking at the results in the previous section, it is striking that the standard deviation of LAI estimation from radar is already large at small LAI values. This is quite contrary to the situation in the optical domain as described before. The comparison between standard

(6)

Similarly as with the CLAIR model, we find an exponential relationship between remote sensing measurement and LAI. Again, the accuracy of the LAI estimate decreases with increasing LAI value. Since we are dealing again with a nonlinear function, a first-order approximation of the standard deviation of the LAI can be derived as

Figure 2. Relation between estimated LAI using the Cloud model and measured LAI for sugar beet in L-band HH polarization (a) and C-band VV polarization (b). Flevoland test site, MAC Europe 1991 campaign. 3 - ooooo L-band HH-pgfl. __ LAl+2*sigma / _~2--°LAI-2*sig/

/

/

./ / ./

a(LAI) = cos 0 / (K'D') • exp(D'" LAI / cos 0)'a(7 ).

/ /

(7)

Because optical remote sensing data can be used for estimating LAI with quite a high accuracy once calibrated, ground measurements of LAI can be replaced by LAI estimates from optical remote sensing data. Since no LAI measurements were performed during MAC Europe 1991 in Flevoland, (optical) AVIRIS data were used to estimate LAI for calibrating the Cloud model. All sugar beet fields, which were present in the AVIRIS optical image as well as in the AIRSAR radar image of the beginning of July 1991, were selected. Data extraction resulted in a total of 37 sugar beet fields for two polarizations (HH, VV) and three frequencies of the AIRSAR (C-, L-, and P-band; resp. 5.3 GHz, 1.25 GHz, and 0.44 GHz). Calibration results showed that L-band H H and C-band VV polarizations were useful. They represent also the configuration of the radar satellites ERS-1 and JERS-1. To calibrate the Cloud model for L-band H H and for C-band VV, a random calibration set of 20 fields was selected from the available fields. Calibration results are given in Table 1. Figure 2 plots the estimated LAI using the Cloud model against the "measured" LAI (from AVIRIS) for the calibration set of MAC Europe. In addition, the lines defining 2 standard deviations from the "measured" LAI are shown.

o

° Oo

/

--1

oi.//

:5

/ 0

O

o°Oo i 1 LAI

0

(a)

i 2

,

measured

/ 3- ooooo C-band W-pal. __ LAl+2*sigma ~o __ LAI-2*slgma " / "

/ /

/.

-13

/

f

O

/ / /

°o o

E ~D 00/

/

0

0

_1 o

o (b)

1 ' 2 L~,I m e a s u r e d

'

/

Combined Use of Optical and Microwave Data

L-band HH-pol. 5-

CLAIR Cloud

__

¢-

approach used optical remote sensing measurements to calibrate the Cloud model, which would not have been possible without optical data.

I I

.o_ 4

I

-I-.,

DYNAMICAL MODELING OF CROP GROWTH

/

"y,

/ / /

©

-o5 "O L...

/

c) 2

/ /

r-

/

-"41

/ /

"1

J

./

o

;

(a)

I

I

2

o

i

4

I

6

LAI

C-band VV-pol. 5-

_ _

CLAIR

__

Cloud

/ / /

0 4_ / /

>

/ / ~

-

/ /

~ 2 -

ml

/ /

0 I o

(b)

45

,

i I

.

J

i 2

,

i 5 LAI

.

i 4

i

i 5

i

i 6

Figure 3. Comparison of standard deviations of LAI estimates from optical and radar measurements: a) L-band HH polarization; b) C-band VV polarization.

deviations of LAI estimates from optical and radar measurements for sugar beet is illustrated in Figure 3. This figure clearly illustrates that the accuracy of LAI estimation from radar measurements is much worse than from optical measurements except for very low LAI values. So, only little additional value is to be expected from radar measurements for LAI estimation when optical measurements are available. The significance of radar measurements is that they can be obtained when optical measurement cannot (mainly caused by bad weather conditions) and may contain information about plant structure (van Leeuwen et al., 1994). Therefore, in the rest of this article emphasis is put on monitoring the growth of crops using growth models when contemporaneous optical and radar measurements are unavailable. However, it must be noted that the above-described contemporaneous

Crop Growth Models Since the 19th century, agricultural researchers have used modeling as a tool to describe relationships between crop growth and yield and environmental factors such as solar irradiation, temperature, and water and nutrient availability. The models compute the daily growth and development rate of a crop, simulating the dry matter production from emergence until maturity. Finally, a simulation of yield at harvest time is obtained. The basis for the calculations of dry matter production is the rate of gross CO2 assimilation of the canopy. Input data requirements include crop physiological characteristics, site characteristics, environmental characteristics, and the initial conditions defined by the date at which the crop emerges. SUCROS (simplified and universal crop growth simulator) (Spitters et al., 1989) is a mechanistic crop growth model that describes the potential growth of a crop from irradiation, air temperature, and crop characteristics. Potential growth means the accumulation of dry matter under ample supply of water and nutrients, in an environment that is free from pests and diseases. The light profile within a crop canopy is computed on the basis of the LAI and the extinction coefficient. At selected times during the day and at selected depths within the canopy, photosynthesis is calculated from the photosynthesis-light response of individual leaves. Integration over the canopy layers and over time within the day gives the daily assimilation rate of the crop (partly from Spitters et al., 1989). The excess assimilate over that needed to maintain the present biomass (maintenance respiration) is converted into new, structural plant matter (with loss due to growth respiration). The newly formed dry matter is partitioned to the various plant organs, through partitioning factors introduced as a function of the phenological development stage of the crop. An important variable that is similated is the LAI, since the increase in leaf area contributes to next day's light interception and hence to next day's rate of assimilation. When applied to operational uses such as yield estimation, models such as SUCROS often appear to fail when growing conditions are nonoptimal (e.g., pest and disease incidence, severe drought, frost damage). Therefore, for yield estimation, it is necessary to "check" modeling results with some sort of information on the actual status of the crop throughout the growing season (Bouman, 1991). Remote sensing can provide such information.

46 Clevers and van Leeuwen

Combination Method Using Inverse Models Clevers et al. (1994) and van Leeuwen and Clevers (1994) described a method for calibrating crop growth models on periodic remote sensing measurements. This was based on a procedure developed by Bouman (1992a) in which remote sensing models were linked to crop growth models so that, for example, canopy reflectance was simulated together with crop growth. The combined growth/reflectance model simulates crop variables, such as biomass and LAI, together with canopy reflectance and WDVI during the growing season. This model now may be calibrated within the biological plausible ranges of parameters used in the growth model to fit the simulated remote sensing signals (WDVI) to the measured ones. Because of the redundancy in the effect of parameter changes in the growth model, the number of parameters to be calibrated in the growth model was kept to a minimum (Clevers et al,, 1994). The parameters "sowing date," "relative growth rate," "light use efficiency," and "maximum leaf area" were selected. In order to use both optical and radar data simultaneously for calibrating the growth model in this study, the growth model was calibrated using the LAI as the essential link. This means that the SUCROS crop growth model is initialized and calibrated to fit simulated LAI values to estimated LAI values obtained from remote sensing measurements. So, the remote sensing models are used in an inverse way (inverse modeling). Thus, first the CLAIR and/or inverted Cloud model are applied for obtaining LAI estimates from the remote sensing measurements. Subsequently, the SUCROS model is calibrated on these LAI estimates. Since we have seen that the accuracy of the LAI estimates depends on the absolute value of the LAI, the reciprocal of the standard deviation of LAI estimation is used as a weighting factor for each individual LAI estimate used in the optimization procedure. To obtain LAI estimates from optical measurements, Eq. (2) is used, and for LAI estimates from radar measurements Eq. (6) is used. In addition, parameter estimates obtained for sugar beet during the calibration of the CLAIR and Cloud model, respectively, are used in these equations. This approach yields at the same time a proper mutual weighting between optical and radar data when data from both windows are used together in the optimization procedure. Moreover, it is obvious that for the methodology it is not relevant whether one has optical and radar data at the same date or not. The applied methodology is illustrated in Figure 4.

MAC EUROPE CAMPAIGN--FLEVOLAND TEST SITE The above will be illustrated with results from the European multisensor airborne campaign MAC Europe in 1991. A description of the MAC Europe campaign for

~g~mOtesensing/

/

servations round/ air /

Field+ weather/ Observations /

/~

+

+

RS model

Crop growth model Crop/soil

optical/microwave

e.g.CLAIR/ Cloud

I_

~

Boundary Conditions

Inversion

~_

of RS models

/

variableranges & initialization e.g.SBFLEVO

| Combination method

Calibration of cropgrowthmodelby RSretrievedvariables

~ffi Yield / level)/ eld

weighted by their accuracy of estimation

Figure 4. Schematic illustration of the methodology for combining remote sensing derived information with crop growth models for yield prediction, as applied in this study.

the Dutch test site Flevoland and of the collected remote sensing and ground truth data was given by Clevers et al. (1994). The most important items will be repeated here. Bouman (1992b) has described a version of the growth model SUCROS that has been initialized for Flevoland. This version is called SBFLEVO and was used in this study. Test Site The test site was located in Southern Flevoland in The Netherlands, an agricultural area with very homogeneous soils reclaimed from the lake IJsselmeer in 1966. The test site comprised 10 different agricultural farms, each 45-60 ha in size. Main crops were sugar beet, potato, and winter wheat. Due to hailstorms and nightfrost damage of the sugar beet in April 1991, some of the sugar beet fields were sown for a second time in late April resulting in large growth differences among fields, Ground Truth Crop parameters concerning acreage, variety, planting date, emergence date, fertilization, harvest date, yield, and occurring anomalies were collected for the main crops. During the growing season, additional parameters were measured in the field approximately every 10 days. The selected parameters were the estimated soil cover by the canopy, the mean crop height, row distance, plants per m 2, the soil moisture condition, and comments about plant development stage.

Combined Use of Optical and Microwave Data 4 7

Meteorological Data Daily meteorological data are needed as input for crop growth simulation models. For the 1991 growing season these were obtained from the Royal Dutch Meteorological Service (KNMI) for the station Lelystad. Data consisted of daily minimum and maximum temperature, daily global irradiation, and daily precipitation. Cropscan Grotmdbased Reflectances Field reflectance measurements were obtained during the 1991 growing season with a ground-based Cropscan radiometer approximately every 10 days. Eight narrowband interference filters with photodiodes were oriented upwards to detect hemispherical incident radiation, and a matched set of interference filters with photodiodes was oriented downwards to detect reflected radiation. Spectral bands were located at 490 nm, 550 nm, 670 nm, 700 nm, 740 nm, 780 nm, 870 nm, and 1090 nm with a bandwidth of 10 nm. The sensor head of the radiometer was mounted on top of a long metal pole and positioned 3 m above the ground surface. The distance to the crop was 2.5-1.5 m, depending on the crop height. As the diameter of the field of view (FOV 28 °) was half the distance between sensor and measured surface, the field of view varied from 1.23 m 2 to 0.44 m 2' CAESAR The CAESAR (CCD Airborne Experimental Scanner for Applications in Remote Sensing) applies linear CCD arrays as detectors. It has a modular setup, and it combines the possibilities of a high spectral resolution with a high spatial resolution. For land applications three spectral bands are available in the green, red, and NIR part of the EM spectrum. Successful overflights over the test site were carried out on 4 July, 23 July, and 29 August 1991. In this study, a spatial resolution of 5 m was applied due to the large fields in the test area. Calibration to reflectances was performed by using artificial reference targets in the field. AVIRIS The ER-2 aircraft of NASA, carrying the Airborne VisibleInfrared Imaging Spectrometer (AVIRIS), performed a successful overflight over the Flevoland test site on 5 July 1991. AVIRIS acquires 224 contiguous spectral bands from 0.41/~m to 2.45/~m. The ground resolution is 20 m as it is flown at 20 km altitude. Calibration to reflectances was performed by the Jet Propulsion Laboratory (JPL) with a new version of the LOWTRAN 7 atmospheric model (van den Bosch and Alley, 1990). AIRSAR The DC-8 aircraft of NASA, carrying the three frequency, full polarimetric AIRSAR of the JPL performed

three successful overflights over the Flevoland test site on 3, 12, and 28 July 1991. The AIRSAR is a synthetic aperture radar that provides P (0.44 GHz), L (1.25 GHz), and C (5.3 GHz) band images at the same time for HH, HV, VH, and VV polarizations. The incidence angle ranges from 20 ° to 60 ° . The standard product is the multilook image (16-look compressed Stokes matrix; cf. Zebker and Lou, 1990). Each pixel represents 6.66 m (range) by 12.1 m (azimuth) in slant range (from 8 m to 12 m in ground range). Calibration was performed by using corner reflectors in the field.

RESULTS OF MAC EUROPE CAMPAIGN 1991 FOR SUGAR BEET Optical Remote Sensing As a reference, first, SBFLEVO was calibrated so that simulated LAI matched LAI values estimated from WDVI measurements performed with the field radiometer for 10 sugar beet fields (Flevoland test site). The radiometer measurements consisted of about 10 measurement dates spread all over the growing season (approximately every 10 days). LAI was estimated using Eq. (2), and Eq. (3) offered an estimate of the accuracy of the LAI estimates, which was used as a weighting factor in the calibration procedure. Results are given in Table 2. The measurements obtained from three CAESAR recordings (4 July, 23 July, and 29 August) during the MAC Europe campaign in 1991 over the Flevoland test area were used for testing the calibration procedure for sugar beet using optical data only. For each date the LAI values for 10 sugar beet fields were estimated from the CAESAR recordings. Subsequently, SBFLEVO was calibrated on these three LAI estimates. Results are given in Table 2. The comparison between estimated and actual (measured) yield is given in Figure 5. Results using only three dates during the growing season in the calibration procedure seem to offer quite satisfactory results, which are not much worse than using the optical radiometer data. On the average, the error of simulated (fresh) beet yield (meaning the average difference between actual and simulated yield) was 4.2 tons/ha (5.5% error) with SBFLEVO calibrated on three CAESAR dates (see Table 2). Radar Remote Sensing From the MAC Europe campaign 1991 three usable AIRSAR recording dates (3, 12, and 28 July) were available for the Flevoland test site. The radar backscatter values of sugar beet fields on 28 July were lower than those on 12 July, which was attributed to plant structural effects (van Leeuwen et al., 1994). Therefore, data from 28 July were left out in this study. The Cloud model was successfully calibrated on the radar data of sugar beet fields from 3 July for both L-band HH polarization

48 Clevers and van Leeuwen

Table 2. Optical and Radar Remote Sensing Configurations with Frequent (10) and Less Frequent (2-5) Observations, Used for the Combination Method and the Accompanying Results, Represented by Yield Errors (Meaning the Average Difference between Actual and Simulated Yield) in Tons per Hectare Remote Sensing Data

Average Error (tons/ha)

Without remote sensing Cropscan (10 dates) CAESAR(3 dates) AIRSARL-HH (2 dates) AIRSARC-VV(2 dates) AIRSARL-HH(2 dates) + CAESAR(3 dates) AIRSARC-VV(2 dates) + CAESAR(3 dates)

13.4 3.7 4.2 9.2 7.2 3.0 3.5

and C-band VV polarization. It was found that the parameters of the Cloud model, as given in Table 1, for both L-band HH polarization and C-band VV polarization, respectively, may also be applied to the measurements of 12 July. As a result, we have two data points during the growing season for a model-based approach using only radar data. By applying Eq. (6) with the appropriate parameter estimates from Table 1, the LAI can be estimated for all sugar beet fields present in both AIRSAR images. Equation (7) offers an estimate of the accuracy of these LAI estimates. Subsequently, SBFLEVO was calibrated on these LAI estimates from the AIRSAR recordings of 3 and 12 July 1991 for the beet fields used before, as far as the corresponding fields were present in both AIRSAR images. Results are given in Table 2 for L-band HH and C-band VV. The comparison between estimated and actual yield is given in Figure 6. Since we have two recording dates rather early in the growing season, accurate yield estimates cannot be expected. On the average, the error of simulated (fresh) beet yield was 9.2 tons/ha (13.0% error) for L-band HH and 7.2 tons/ha (9..8% error) for C-band VV, respectively, with SBFLEVO calibrated on two AIRSAR dates. This is better than the result obtained with "standard" SBFLEVO without remote sensing information (which gave an average error of simulated fresh beet yield of 13.4 tons/ha; see also Table 2). Moreover, in this case C-band VV polarization offers better results than L-band HH polarization. For sugar beet this is about the best we can expect using only the model-based approach on radar data, since after mid-July (in 1991, The Netherlands) the Cloud model cannot be applied anymore, because of saturation of the backscatter after closure of the crop (van Leeuwen, 1996). The Combination of 3 Optical and 2 Radar Observations In this section, LAI estimates from the three CAESAR recordings and the two AIRSAR recordings are integrated and, with their appropriate weighting factors, used for calibrating SBFLEVO. Results are given in Table 2 for the three CAESAR recordings in combina-

tion with L-band HH and C-band VV radar data. The comparison between estimated and actual yield is given in Figure 7. On the average, the error of simulated (fresh) beet yield was 3.0 tons/ha (4.2% error) for L-band HH and 3.5 tons/ha (4.8% error) for C-band VV, respectively. This error is clearly smaller than the one obtained for the three CAESAR recording dates only and of the same magnitude as the errors obtained using the Cropscan (optical) measurements during the whole growing season. These results indicate a synergistic effect by using both optical and radar data for crop growth monitoring. However, under practical conditions optical data will be available for only a few dates during the growing season. For instance, had no optical data from 4 July been available, radar data for early July would have significantly improved the monitoring of crop growth early in the growing season. As mentioned before, another potential advantage of radar measurements lies in the possibility of obtaining information about crop structure changes. The latter may be related to important transitions in crop development stage (van Leeuwen et al., 1994; van Leeuwen,

DISCUSSION AND CONCLUSIONS For simultaneous (contemporaneous) observations no synergy occurred in the estimation of LAI. Optical data were most suitable. Calibration of the Cloud model at one date (contemporaneous) is possible using optical data ff enough fields are available for the calibration and the among-field variation is large. For operational applications the assumption of nonsimultaneous observations is most realistic. For sugar beet, radar data can only be used for estimating LAI early in the growing season (before crop closure). This may be called a model-based approach. After crop closure, radar backscatter is determined by crop architecture (leaf angle distribution). However, this still may yield important information for crop growth monitoring. Using the latter information may be called a featurebased approach. Results for sugar beet indicated that, when periodic

Combined Use of Optical and Microwave Data 49

estimated beet yield 1 0 0 - (tons/ha)

estimated beet yield 1 0 0 - (tons/ha) 90,

90

80

8O 0 0

70

70

60

6O

50

D

(a)

I

60

50

[

I

l

70

I

~

I

80

'

90

I

50

100

actual beet yield (tons/ha)

(tons/ha)

'

I

70

I

80

'

I

'

90

I

100

yield (tons/ha)

(tons/ha)

9O

80-

o

70.

70

0

,

0 1

,

I

'

I

50

i

60

70

80

90

QO

6 0

'

,

50

(b)

'

actual beet

100

80-

50

In

60

estimated beet yield

9O

6

' 50

(a)

estimated beet yield 100

o o

100

actual beet yield (tons/ha)

m)

: 50

0 jO I I I I ' 60

actual beet

.

i

70

80

90

1 O0

yield (tons/ha)

Figure 5. Comparison between estimated yield and actual yield for 10 optical groundbased Cropscan measurements (a) and 3 aerial optical (CAESAR) recording dates (b). The "ideal" h l line is also depicted.

Figure 6. Comparison between estimated yield and actual yield for two AIRSAR recording dates in L-band HH polarization (a) and C-band VV polarization (b). The "ideal" 1:1 line is also depicted.

optical recordings are available, LAI can be monitored well and a good estimate of sugar beet yield at the end of the season is possible by using a calibrated crop growth model. When only a few recording dates with an optical sensor are available, radar recordings at L-band HH polarization or C-band VV polarization improved the results of crop monitoring and yield estimation compared with the optical data only. This confirms that

the main advantage of radar lies in acquiring information on crop growth when other techniques (in particular optical techniques) fail. Different scenarios with various combinations of optical and radar data at various dates during the growing season for sugar beet still must be evaluated further in order to obtain an accurate picture of the significance of radar data (in a model-based and feature-based ap-

50 Clevers and van Leeuwen

estimated beet yield (tons/ha) 100-

combined use of optical and radar measurements, the translation to the use of spaceborne data must be studied. The role of optical satellites like Landsat-TM and SPOT must be evaluated in relation to the role of radar satellites like ERS-1 and JERS-1.

90. This article describes a study carried out within the framework of the NRSP-2 under responsibility of the Netherlands Remote Sensing Board (BCRS) and Contract No. 9 8 3 7 / 9 2 / N L / G S of ESA. Moreover, we thank NASA and ESA for providing the A VIRIS and AIRSAR data during the framework of MAC Europe 1991.

80 ¸ o

°o

70 o

REFERENCES

60

5O

I

5O

I

I

60

I

I

70

I

I

80

I

t

90

I

100

actual beet yield (tons/ha)

(a)

estimated beet yield

(tons/ha)

100

80 70 60

~oo

50

I

50

(b)

~o

I

60

I

I

70

I

I

80

I

I

90

'

I

100

actual beet yield (tons/ha)

Figure 7. Comparison between estimated yield and actual yield for three CAESAR recording dates and for two AIRSAR recordings in L-band HH polarization (a) and C-band VV polarization (b). The "ideal" h l line is also depicted.

proach) for crop growth monitoring. The technique to calibrate a crop growth model with remote sensing data was developed for sugar beet in Flevoland. The behavior of other crops (other varieties, other locations) can be quite different. The developed techniques, therefore, have to be evaluated for other cases (crops, areas) as well, which eventually may result in a more generalized approach. Moreover, for operational applications of the

Attema, E. P. W., and Ulaby, F. T. (1978), Vegetation modelled as a water cloud, Radio Sci. 13:357-364. Bouman, B. A. M. (1991), Linking X-band radar backscattering and optical reflectance with crop growth models, Ph.D. thesis, Wageningen Agricultural University, Wageningen, The Netherlands. Bouman, B. A. M. (1992a), Linking physical remote sensing models with crop growth simulation models, applied for sugar beet, Int. J. Remote Sens. 13:2565-2581. Bouman, B. A. M. (1992b), SBFLEVO and WWFLEVO, growth models to simulate crop growth, optical reflectance and radar backscatter of sugar beet and winter wheat, calibrated for Flevoland, CABO-DLO Report 163, Wageningen, The Netherlands, 61 pp. plus appendices. Bouman, B. A. M., van Kasteren, H. W. J., and Uenk, D. (1992), Standard relations to estimate ground cover and LAI of agricultural crops from reflectance measurements, Eur. J. Agron. 4:249-262. Clevers, J. G. P. W. (1988), The derivation of a simplified reflectance model for the estimation of Leaf Area Index, Remote Sens. Environ. 25:53-69. Clevers, J. G. P. W. (1989), The application of a weighted infrared-red vegetation index for estimating Leaf Area Index by correcting for soil moisture, Remote Sens. Environ. 29:25-37. Clevers, J. G. P. W., Bilker, C., van Leeuwen, H. J. C., and Bouman, B. A. M. (1994), A framework for minitoring crop growth by combining directional and spectral remote sensing information, Remote Sens. Environ. 50:161-170. Del~colle, R., Maas, S. J., Gu~rif, M., and Baret, F. (1992), Remote sensing and crop production models: present trends, ISPRS J. Photogramm. Remote Sens. 47:145-161. Hoekman, D. H., Krul, L., and Attema, E. P. W. (1982), A multilayer model for radar backscatter from vegetation canopies, in Proc. IGARSS'82 Symposium, Munich, IEEE New York, Volume II, pp. 4.1-4.7. Maas, S. J. (1988), Use of remotely sensed information in agricultural crop growth models, Ecol. Model. 41:247-268. Spitters, C. J. T., van Keulen, H., and van Kraailingen, D. W. G. (1989), A simple and universal crop growth simulator: SUCROS87, in Simulation and Systems Management in Crop Protection (R. Rabbinge, S. A. Ward, and H. H. van Laar, Eds.), Simulation Monographs 32, Pudoc, Wageningen, pp. 147-181.

Combined Use of Optical and Microwave Data 51

van den Bosch, J., and Alley, R. (1990), Application of LOWTRAN 7 as an atmospheric correction to Airborne Visible / InfraRed Imaging Spectrometer (AVIRIS) data, in Proceedings of the Second Airborne Visible / InfraRed Imaging Spectrometer (AVIRIS) Workshop (R.O. Green, Ed.), JPLPublication 90-54, Pasadena, CA, pp. 78-81. van Leeuwen, H. J. C. (1992), Multifrequency and multitemporal analysis of scatterometer radar data with respect to agricultural crops using the Cloud model, in Proc. MAESTRO /AGRISCATT Final Workshop, ESTEC, ESA wpp31, ESA, Paris, pp. 203-207. van Leeuwen, H. J. C. (1996), Methodology for combining optical and microwave remote sensing in agricultural crop growth monitoring, Ph.D. thesis, Wageningen Agricultural University, Wageningen, The Netherlands.

van Leeuwen, H. J. C., and Clevers, J. G. P. W. (1994), Synergy between optical and microwave remote sensing for crop growth monitoring, in Proc. Sixth International Symposium Physical Measurements and Signatures in Remote Sensing, Val d'Is~re, France, CNES, Paris, pp~ 11751182. van Leeuwen, H. J. C., Clevers, J. G. P. W., and Rijckenberg, G. J. (1994), Synergetic use of optical and microwave remote sensing data using models and specific features with respect to the sugar beet crop, in Proc. IGARSS 94 Symposium, Pasadena, California, IEEE, Piscataway, NJ, Vol. II, pp. 827-831. Zebker, H. A., and Lou, Y. (1990), Phase calibration of imaging radar polarimeter Stokes matrices, IEEE Trans. Geosci. Remote Sens. 28:246-252.