Ph. D. Dissertation defense Evaluation of the Performance of the MODIS LAI and FPAR Algorithm with Multiresolution Satellite Data Yuhong Tian Department of Geography, Boston University Dissertation Committee Ranga B. Myneni Yuri Knyazikhin Mark A. Friedl Curtis E. Woodcock Alexander L. Marshak 1 of 39

Contents 1. Introduction 2. Objectives 3. Research Topics  Prototyping of the MODIS LAI/FPAR algorithm with LASUR and Landsat data  Radiative transfer based scaling of LAI/FPAR retrievals from reflectance data of different resolutions  Multiscale analysis and validation of the MODIS LAI over Maun, Botswana 4. Concluding Remarks 5. Future Work 2 of 39

1. Introduction LAI and FPAR: Definition LAI: green leaf area index, one-sided green leaf area per unit ground area. FPAR: fraction of photosynthetically active radiation (  m) absorbed by the vegetation. Importance They are key variables in land surface models for calculation of surface photosynthesis, evapotranspiration, and net primary production. 3 of 39

2. Objectives  Prototyping: 1.To test the physical functionality and performance of the algorithm with MODIS like data.  Effects of spatial resolution: 1.To inquire about the cause of the discrepancy between coarse and moderate resolution output from the same algorithm. 2.To investigate the adjustment of retrieval techniques for data resolution.  Validation: 1.To derive uncertainty information on LAI/FPAR product by comparing with field data. 2.To define a sampling strategy. 4 of 39

3. Research Topics Part One Prototyping of the MODIS LAI/FPAR Algorithm with LASUR and Landsat data Tian et al., “Prototyping of MODIS LAI and FPAR algorithm with LASUR and Landsat data”. IEEE Trans. Geosci. Remote Sens. 38(5): , of 39

Research Topics: Part one Introduction Prototyping: use data from other instruments to test the functionality of the MODIS algorithm before MODIS data are available. 6 of 39

Research Topics: Part one Data Land surface reflectances (LASUR). - spatial resolution: 1/7th of a degree - RED ( nm) and NIR ( nm) from July 1989 A TM image of Northwest U.S. (Washington and Oregon) m resolution - RED ( nm) and NIR ( nm) from June 26, A biome classification map (BCM). - Grasses and cereal crops - Shrubs - Broadleaf crops - Savannas - Broadleaf forests - Needle forests 7 of 39

Research Topics: Part one Consistent with Physics 8 of 39

Research Topics: Part one Impact of Biome Misclassification Misclassified Biome Type Grasses and Cereal Crops Shrubs Broadleaf Crops Savannas Broadleaf Forest Needle Forests BCM Biome Type Grasses and Cereal Crops Shrubs Broadleaf Crops Savannas Small effectLarge effect Broadleaf Forest Needle Forests Large effect Small effect 9 of 39

Research Topics: Part one Scale Dependence of the Algorithm LANDSAT Fine resolution LASUR Coarse resolution 10 of 39

Research Topics: Part one Conclusions Prototyping results demonstrate the ability of the algorithm to produce global LAI and FPAR fields. The LAI and FPAR fields follow regularities expected from physics. The algorithm is dependent on the spatial resolution of the data. 11 of 39

Research Topics Part Two Radiative transfer based scaling of LAI/FPAR retrievals from reflectance data of different resolutions Tian, et al., “Radiative transfer based scaling of LAI/FPAR retrievals from reflectance data of different resolutions”. Remote Sens. Environ., 2001 (in review). 12 of 39

Research Topics: Part two Introduction The goal of scaling: values of LAI derived from coarse resolution sensor data should equal the arithmetic average of values derived independently from fine resolution sensor data. Scaling issues arise when one attempts to assemble a consistent time series of LAI/FPAR products with data from different spatial resolutions. one attempts to validate moderate resolution (~ 1 km) sensor products with field measurements at much finer resolutions. 13 of 39

Research Topics: Part two Objectives To investigate the effect of pixel heterogeneity on LAI/FPAR retrievals. To develop a physically based theory for scaling with scale dependent radiative transfer formulation. 14 of 39

Research Topics: Part two Data AVHRR land surface reflectances at 1 km resolution over North America for July RED ( nm) and NIR ( nm). - 1 km AVHRR reflectance data were aggregated to 4, 8, 16, 32 and 64 km resolutions. A six biome map of North America - developed from 1 km AVHRR NDVI data of 1995 and 1996 by Lotsch et al. (2000). 15 of 39

Research Topics: Part two Characterizing Land Cover Heterogeneity Percentage function (pf). B1 B4 B5 B4 B5 B1B5 pf 1 =3/16; pf 4 =4/16; pf 5 =9/16 Purity of this pixel=pf 5 =9/16 A 4 km x 4 km resolution pixel 16 of 39 The percentage occupation of subpixel biome type in a given coarse resolution pixel. “purity” of a pixel. Percentage function of dominant biome type in a given coarse resolution pixel.

Research Topics: Part two Purity Decreases as Spatial Resolution Decreases Resolution (km) Biome purity > 90% Percentage of pixels Biome purity < 50% 17 of 39

Research Topics: Part two Purity Has a Strong Effect on LAI Retrievals Error = |LAI true -LAI estimated |/LAI true 18 of 39 B1 B4 B5 B4 B5 B1B5 Dominant Land Cover Purity a)Reflectance at 4 km resolution b)Biome type Input:

Research Topics: Part two Energy conservation law as a tool to scale models 19 of 39 Leaf absorption: (1-  )N oi Leaf scattering:  N oi Leaf interception: p  N oi N oi N ot (1-p)  N 0i N i = N oi + p  N oi + (p  ) 2 N oi +… = N oi /(1-p  ) N t = N ot /(1-p t  )  = 1/(1-pf soil )  j pf j N N = N R + N t + (1-  )N oi  : leaf albedo, the portion of radiation flux density incident on the leaf surface that the leaf transmits and reflects. p: the fraction of photons that are scattered by leaves and will interact with leaves again.

Research Topics: Part two 20 of 39 Dominant Land Cover Purity Improved Retrieval Accuracy BeforeAfter Dominant Land Cover Purity

Research Topics: Part two Conclusions LAI retrieval errors are inversely related to the proportion of the dominant land cover in a coarse resolution pixel. Pixel heterogeneity must be accounted to improve accuracy in retrievals. 21 of 39

Research Topics Part Three Multiscale analysis and validation of the MODIS LAI over Maun, Botswana Tian, et al., “Multiscale analysis and validation of the MODIS LAI over Maun, Botswana”. Remote Sens. Environ., 2001 (submitted in October, 2001). 22 of 39

Research Topics: Part three Introduction As MODIS LAI and FPAR data start to become publicly available, product quality must be ensured by validation. Validation: the process of assessing the uncertainty of data products by comparison to reference data (e.g., in situ, aircraft, and high-resolution satellite sensor data). 23 of 39

Research Topics: Part three Objectives To develop an appropriate ground-based validation technique for assessing the uncertainties in MODIS LAI product. To develop sampling strategies to collect data needed for validation of the MODIS LAI product. 24 of 39

Research Topics: Part three Data LAI measured by LAI-2000 Plant Canopy Analyzer. Landsat ETM+ (30 m) data. MODIS reflectance data (1 km) simulated from ETM+. 25 of 39

Research Topics: Part three Pandamatenga S 18  39.5 E 25  29.8 Maun S 19  55.8 E 23  30.7 Okwa S 22  24.6 E 21  42.8 Tshane S 24  10.1 E 21  of 39

Sampling Scheme 1000 m N375W 0 N375E A375W A375E B375W B375E 750 m 250 m 300 m 250m A C D E F B START POINT 25m N Research Topics: Part three 27 of 39

Research Topics: Part three Validation of the MODIS LAI Product At Maun Field dataETM+MODIS product Problems with validation Only four pairs of pixels between field measurements and MODIS data. Spatial registration is not accurate. 28 of 39

Research Topics: Part three ETM+ ImageSegmentation Map Patch by Patch Comparison Shortcomings of pixel by pixel comparison GPS readings are not accurate. Measured LAI values have high variation over short distances. 29 of 39

Research Topics: Part three Consistency between LAI Retrievals and Field Measurements 30 of 39 s LAI-Field Measurements LAI-Algorithm Retrievals

Research Topics: Part three Underestimation of LAI for Coarse Resolution Data 31 of 39

Research Topics: Part three D1D1 D2D2 D3D3 Hierarchical Analysis of Multiscale Variation in LAI Data Four scale levels: whole image > class > region > pixel D 1 =D 11 + D 12 +D 13 D 2 =D 21 + D 22 D 3 =D 31 + D 32 +D 33 D 11 D 12 D 13 D 31 D 32 D 33 D 22 D of 39 Four images: image effect, class effect, region effect, pixel effect

Research Topics: Part three Distance (h) Semivariance (  ) sill range Semivariogram Analysis for 4 Scale Levels 33 of 39

Three Sites Maun (Botswana) Research Topics: Part three Harvard Forest (USA) Ruokolahti Forest (Finland) 34 of 39

LAI Semivariograms Maun Research Topics: Part three Harvard ForestRuokolahti Forest 35 of 39 Pixel Effect Region Effect Class Effect Original Image Semivariance

Research Topics: Part three Conclusions Consistency between LAI retrievals from 30 m ETM+ data and field measurements indicates satisfactory performance of the algorithm. Hierarchical variance analysis shows that the LAI retrievals from ETM+ data demonstrate multiple characteristic scales of spatial variation. 1.Within the three sites, patterns of variance in the class, region, and pixel scale are different with respect to the importance of the three levels of landscape organization. 2.The spatial structure is small across the three sites. Validation needs to be performed over small areas. 3.For validation activities, patches are better than individual pixels unless sample and registration accuracy are outstanding. 36 of 39

4. Concluding Remarks Prototyping results demonstrate the ability of the algorithm to produce global LAI and FPAR fields. The LAI and FPAR fields follow regularities expected from physics. LAI retrieval errors are inversely related to the proportion of the dominant land cover in a coarse resolution pixel. A physically based theory for scaling with a scale dependent radiative transfer formulation was developed. 37 of 39

Consistency between LAI retrievals from 30 m Landsat ETM+ data and field measurements from Maun (Botswana) indicates satisfactory performance of the algorithm. LAI fields demonstrate multiple characteristic scales of spatial variation. Isolating the effects associated with different scales through variograms aids the development of a new sampling strategy for validation of MODIS products. 38 of 39

5. Future Work Use MODIS products to improve representation of the land surface in global climate models using the scaling ideas developed here. 39 of 39

The end

Questions?

The MODIS LAI/FPAR algorithm r k : modeled BRDF d k : satellite measured BRDF  k : uncertainties in measurements and simulations p=[canopy, soil] LAI Frequency Solution distribution function

Research Topics: Part one Retrieval Index Depends on Quality of Surface Reflectance Uncertainty Retrieval Index: ratio of the number of retrieved pixels to total number of pixels.

X ijk = I + C i + R ij + P ijk I: the image effect; C i : the effect associated with class i; R ij : the effect associated with region j of class i; P ijk : the pixel effect associated with pixel k of region j of class i. Research Topics: Part three I =  (D) C i =  (D i ) -  (D) R ij =  (D ij ) -  (D i ) P ijk = X ijk -  (D ij ) Image decomposition