ESTIMATING LEAF AREA INDEX BY COUPLING RADIATIVE TRANSFER MODEL AND A DYNAMIC MODEL FROM MULTI-SOURCE REMOTE SENSING DATA

Xiyan Wu†, Zhiqiang Xiao*† Jindi Wang†

† State Key Laboratory of Remote Sensing Science, Jointly Sponsored by Beijing Normal University and Institute of Remote Sensing Applications, CAS

Department of Geography and Remote Sensing, Beijing Normal University, China, 100875 *Corresponding author. Email: zhqxiao@bnu.edu.cn

ABSTRACT

Satellite remote sensing enables derivation of LAI globally at available spatial resolution and temporal frequency, and several LAI products have been produced. However, there are problems for the current global or regional LAI products, which restrict the application of these products. On the one hand, there are gaps between the large number of parameters of physical models and the small amount of data obtained by single sensor, which may cause the decrease in accuracy that LAI products should have. On the other hand, there are gaps between instantaneous observation of remote sensing and parameters which have change rules. Custom methods to retrieve LAI from remote sensing data are just involving the transient observations, discarding the information about process. To resolve these problems, we develop a methodology to retrieve LAI by involving diverse data from multiple sensor and LAI change rule. The methodology can take full advantage of the different band and angle information of time series MODIS and MISR data to improve the accuracy of the retrieved LAI over the MODIS LAI product compared to the field measured LAI data. And the retrieved LAI is also temporally continuous.

Index Terms—MISR, MODIS, LAI, Assimilation, Coupled model

1. INTRODUCTION

Leaf area index (LAI) is an important parameter in canopy interception, evapotranspiration, and net photosynthesis. So it’s an important research topic to retrieve accurate LAI in the field of remote sensing. Many efforts have been made to estimate LAI from satellite measurements through diverse methods. The most well-known methods are relationship between Vegetaion Indexs and LAI [1] and physical model inversion [2], which are widely used in this field. Support vector machines regression [3], Neural Network [4], and other method is available to retrieve LAI. Some methods, taking advantage of multi-source measurements [5, 6], were developed in the past several years. However, most of these

methods are based on the transient remote sensing observations and the retrieved results are temporally discontinuous. These temporally discontinuous LAI can not satisfy the requirement of applications, which demands continuous and more accurate LAI. In recent years, data assimilation method is introduced to estimate LAI, which can solve the problem above. Some researches have been performed to analyze the capability of data assimilation method [7], and it becomes a concerned method in this field. Xiao et al. [8] developed a temporally integrated inversion method to estimate LAI from time series MODIS reflectance data. Results show that the algorithm is able to produce temporally continuous LAI product, and the accuracy has been significantly improved over the MODIS LAI product compared to the field measured LAI data. However, the algorithm just used MODIS reflectance data. In this paper, we will further analyze the potential of retrieval of LAI from different sensors which may contribute to improving the accuracy of the retrieved LAI. And the MODIS and MISR reflectance observation are used to test the algorithm.

2. DATA

Data from three AmeriFlux sites are chosen for our research in this paper. The attributes of these flux sites are given in Table 1. We use the MODIS collection 5 surface reflectance product with the spatial resolution of 500-m and the temporal resolution of 8-days, and the MISR Level 2 surface product of Bidirectional reflectance factor (BRF) as input data in our methodology. The characteristic of the two types of observation is as followed: (1) MODIS observation data are evenly distributed in time-series, while MISR observation data are distributed less evenly. (2) MISR observation data include multi-angle information while MODIS observation data is single-angle in a day.

Only the red and near infrared channels are used to retrieve LAI for both type of data set. Considering the quality of MODIS reflectance data which may affect by many factors such as clouds, aerosols, water vapor and ozone, the low quality MODIS reflectance data are

III - 658978-1-4244-3395-7/09/$25.00 ©2009 IEEE IGARSS 2009

eliminated. Only the data with good quality is used as input data. Comparatively speaking, MISR BRF data is in better quality, so we do not check the MISR BRF data.

For matching pixels from different observations, we use the GCTP software to transform MISR reflectance data with SOM projection, to SIN projection, used in MODIS data, and resample the MISR reflectance data with a resolution of 1.1km into 1km.

3. ASSIMILATION METHOD

We develop a methodology to retrieve LAI by involving diverse data from multiple sensor and LAI change rule. The algorithm couples the Markov chain reflectance model (MCRM), which is a canopy radiative transfer model, with double logistic model, which is a LAI temporal profile model. LAI is an important input parameter of radiative transfer model, and it is also an output parameter of the LAI temporal profile model. Thus, these two models can be combined through LAI parameter.

A preliminary analysis using time series MODIS and MISR data was performed for validating this method. The cost function to be minimized is as follows:

1

1

1

1

1

( )

[ ( )] [ ( )]

[ ( )] [ ( )]

Tb b

NT

i i i i i i ii

M T

j j j j j j jj

J

H LAI H LAI

H LAI H LAI

B

y R y

y R y

Where N is the observation times of the MODIS sensor in crop growing season, iy is the surface reflectance obtained by MODIS sensor, M is the observation times of the MISR sensor in crop growing season, jy is the surface reflectance obtained by MISR sensor, is the set of input parameters to be estimated, and b is the a priori information on these parameters, LAIi( ) and LAIj( ) is logistic model, H(·) is the Markov chain reflectance model (MCRM), B and R are the background error covariance and the observation error covariance respectively. In this paper, B is set according to land cover type. In consideration of some factors affecting the quality of some observation data, we give the higher weight for the data of higher quality or more information, in

opposite condition, we give a lower weight. So the performance of more accurate observations is enhanced.

We use SCE-UA algorithm, which is a global optimization strategy, to minimize the cost function.

4. RESULTS

As introduced in data section, the three sites are used to test the algorithm. The vegetation type for the Bondville and Mead site is croplands, while the vegetation type for the Tonzi Ranch site is grass.

In this section, we will analysize the results by comparing retrieved LAI with field LAI, MODIS LAI product, and results from traditional methodology, which only use the transient remote sensing observations from one sensor.

At Bondville, the available days for MODIS are about two times of MISR, and cover the most important period in growing season.

The retrieved LAI of Bondville is shown in Figure 1. It is compared with field LAI and LAI product of MODIS. As show in Figure 1 (a), although gaps still exist between retrieved LAI and Field LAI, they are obviously smaller than gaps between LAI product of MODIS and field LAI.

The result is also compared with retrievals from observations of MODIS and MISR in Figure 1 (b). In traditional methodology, retrieved LAI covers only few days in a whole year. Moreover, because some observations of MODIS and MISR are not on the same day, it is hard to combine these two sources of data from different two days. Most methodologies of combining these two sources of data are based upon the assumption that vegetation doesn’t change between the two days. However, because some types of vegetations generally change rapidly in growing season, methodologies based on such assumption has its inherent limitation. The algorithm in this paper can better fill the gaps better than traditional methodology.

Figure 1 (a)

Table 1 Information of each test site Site

Name Land Cover Type Longitude,Latitude

Testingtime period

Bondville Croplands(maize) 88.291867 W, 40.006100 N

2001

Mead Croplands(soybean) 96.439646 W, 41.179667 N

2004

TonziRanch

Oak savanna, grazed grassland

120.96598 W, 38.431600 N

2001-2002

1 33 65 97 129 161 193 225 257 289 321 353

DAY/2001

0

1

2

3

4

5

LAI(

m2 /

m2 )

measured LAIMOD15LAI_MOD+MISRMOD_DAYMISR_DAYMOD MISR_DAY

III - 659

Figure 1 (b) Figure 1 (a) Retrieved LAI at Bondville site compared with field LAI and LAI product of MODIS. “MOD15” represents the LAI product of MODIS, “LAI_MOD+MISR” represents retrieved LAI using the methodology in this paper, “MOD_DAY” marks days of using observations of MODIS, “MISR_DAY” marks days of using observations of MISR, “MOD MISR_DAY” marks days of using both MODIS and MISR.Figure 1 (b) Retrieved LAI at Bondville site compared with retrievals from observations of MODIS and MISR respectively. “LAI_MODp” represents LAI retrieved from one day observation of MODIS, “LAI_MISRp” represents LAI retrieved from one day observations of MISR. “LAI_MOD+MISR” represents retrieved LAI using the methodology in this paper.

Available in growing season ate Mead Site are about twenty days. The available days of MODIS are also about two times of MISR, and cover the most important period in growing season. There are also observations from MISR at the same time, but available days are much fewer than MODIS.

The retrieved LAI at Mead site is show in Figure 2 (a).Although retrieved LAI cannot perform well at the beginning and the most flourishing period of growing season, it also has a more similar trend to field LAI than LAI product of MODIS.

The comparison on retrieved LAI of our methodology and traditional method is show in Figure 2 (b). In traditional method, the change tendency is twisted during growth season. The instability increases markedly between 187 and 260, and the results are so different between MODIS and MISR. On the one side, the imprecision and inconsistency of LAI retrieved from MISR observation is greater than MODIS. On the other side, because retrieved LAI from MISR are in different day with MODIS, the growth tendency seems quite different between them. So it is difficult and inaccurate to retrieve LAI from two data sets, which have obvious floating results. In our algorithm, however, according to the continuous result, these instability and disturbance, as showed in figure 2 (b), are minimized.

At Tonzi Ranch site, the starting time and closing time of growing season are not significantly appointed, so we use the available data in 365 days. The span of interesting period is from 241 of 2001 to 233 of 2002. In 2002, the

number of available dates for MODIS and MISR are approximately equal, and these observations cover the most important growing period.

1 33 65 97 129 161 193 225 257 289 321 353

0

1

2

3

4

5

LAI(

m2 /

m2 )

measured LAILAI_MODpLAI_MISRpLAI_MOD+MISR

1 33 65 97 129 161 193 225 257 289 321 353

DAY/2004

0

1

2

3

4

5

LAI(m

2 /m

2 )

measured LAIMOD15LAI_MOD+MISRMOD_DAYMISR_DAYMOD MISR_DAY

Figure 2 (a)

1 33 65 97 129 161 193 225 257 289 321 353

DAY/2004

0

1

2

3

4

5

LAI(m

2 /m2 )

measured LAILAI_MODpLAI_MISRpLAI_MOD+MISR

Figure 2 (b)

Figure 2 (a) Retrieved LAI at Mead site compared with field LAI and LAI product of MODIS. The symbol is the same as Figure 1 (a).Figure 2 (b) Retrieved LAI at mead site compared with retrievals from observations of MODIS and MISR respectively. The symbol is the same as Figure 1 (b).

The retrieved LAI at Tonzi Ranch site is shown in Figure 3 (a). It is compared with field LAI and LAI product of MODIS in this figure. As show in Figure 3 (a), the values of LAI product of MODIS has been evidently overestimated, six of which are about two times of field LAI. In comparison to LAI product of MODIS, our results more closely approach field LAI, especially before the date of 122 in 2002.

The comparison on retrieved LAI of our methodology and traditional method is show in Figure 3 (b). Retrieval from observation of MODIS and MISR respectively are close to each other at this site. Results are performed well by both of traditional method and our method in this paper before the date of 122 in 2002, and overestimated after the date of 122 in 2002. Change tendency is slightly clearer in the results of our method than traditional method.

III - 660

Figure 3 (a)

Figure 3 (b) Figure 3 (a) Retrieved LAI at Tonzi Ranch site compared with field LAI and LAI product of MODIS. The symbol is the same as Figure 1 (a).Figure 3 (b) Retrieved LAI at Tonzi Ranch site compared with retrievals from observations of MODIS and MISR respectively. The symbol is the same as Figure 1 (b).

5. CONCLUSIONS In this paper, we develop an algorithm which combines observations of MODIS and MISR to retrieve temporally continuous and smooth LAI profile. The time series MODIS surface reflectance data and MISR BRF data at three American flux stations, are used to validate our algorithm. The capability of this inversion algorithm is discussed and compared with the LAI product of MODIS and the retrieved LAI from traditional method. In our algorithm, retrieved LAI is continuous, and significantly improve the accuracy of the retrieved LAI over the MODIS LAI product compared to the field measured LAI data. Besides, two kinds of observations can be combined in our inversion algorithm, though they have different character in time-series distribution and visual angle. On account of the different character in time series distribution, data from different source may on different days, which is a problem in combination. Most algorithms combine multi-sensor observations work under the assumption that there is no

change among different days, consequently error will increases when vegetation changes rapidly in growing season. Our algorithm doesn’t work under the assumption used in these algorithms, and consider the difference in time series. So it will never neglect the rapid change in growing season.

0.0

1.0

2.0

3.0LA

I (m

2 /m2 )

measured LAIMOD15LAI_MOD+MISRMOD_DAYMISR_DAYMOD MISR_DAY

DAY/2001-2002241 281 321 36133 73 113 153 193 233

6. ACKNOWLEDGEMENTS

This research was supported by Chinese 973 Program under grant 2007CB714407, the National Natural Science Foundation of China under grant 40871163, 40701102, National Science and Technology Support Project under grant 2008BAC34B03 and the EU Seventh Framework Programme (CEOP-AEGIS).

0.0

1.0

2.0

3.0

LAI (

m2 /m

2 )

measured LAILAI_MODpLAI_MISRpLAI_MOD+MISR

DAY/2001-2002241 281 321 361 33 73 113 153 193 233

7. REFERENCES

[1] J. Wu, D. Wang and M. Bauer, “Assessing broadband vegetation indices and QuickBird data in estimating leaf area index of corn and potato canopies,” Field Crops Research, pp. 33-42, 2007.

[2] J. Qi, Y. Kerr, M. Moran, M. Weltz, A. Huete, S. Sorooshian and R. Bryant, “Leaf area index estimates using remotely sensed data and BRDF models in a semiarid region,” Remote sensing of environment, pp. 18-30, 2000.

[3] S. Durbha, R. King and N. Younan, “Support vector machines regression for retrieval of leaf area index from multiangle imaging spectroradiometer,” Remote sensing of environment, pp. 348-361, 2007.

[4] H. Fang and S. Liang, “Retrieving leaf area index with a neural network method: simulation and validation,” IEEE Transactions on Geoscience and Remote Sensing, pp. 2052-2062, 2003.

[5] R. Houborg and E. Boegh, “Mapping leaf chlorophyll and leaf area index using inverse and forward canopy reflectance modeling and SPOT reflectance data,” Remote sensing of environment, pp. 186-202, 2008.

[6] H. Fang and S. Liang, “A hybrid inversion method for mapping leaf area index from MODIS data: experiments and application to broadleaf and needleleaf canopies,” Remote sensing of environment, pp. 405-424, 2005.

[7] L. Dente, G. Satalino, F. Mattia and M. Rinaldi, “Assimilation of leaf area index derived from ASAR and MERIS data into CERES-Wheat model to map wheat yield,” Remote sensing of environment, pp. 1395-1407, 2008.

[8] S. L. Z. Xiao, J. Wang, J. Song, X. Wu,, “A temporally integrated inversion method for estimating leaf area index from MODIS data,” IEEE Transactions on Geoscience and Remote Sensing, (In press).

III - 661