*2.7. Field Measured LAI*

OLIVE (On Line Interactive Validation Exercise) platform is established by CEOS-LPV (Committee on Earth Observation System Land Product Validation) and it is devoted to the validation of global remotely sensed land surface products [23]. The OLIVE data has an independent database named DIRECT that include true LAI dataset that are all collected from the existing experimental networks such as FLUXNET, VALERI, Bigfoot, et al., whose sites are all selected at homogenous land cover types and can present 9–100 km<sup>2</sup> spatial range, and have also been utilized in other LAI product validation studies [22,24,29]. The dataset can be downloaded from http://calvalportal.ceos.org/web/olive/ site-description. According to our study time range, we finally selected 47 field LAI measurements over 37 sites from the true LAI database with six biome types; one of them is located in China. And considering our study area, we also found another six field measurements located in China from the present literature [30]. They first established the empirical relationships between clear-sky TM/ETM+ image vegetation index such as NDVI (Normalized Difference Vegetation Index), SR (Simple Ratio), RSR (Reduced Simple Ratio), etc. and field measured LAI in the 30 × 30 m sampling plots, then generated fine-resolution LAI maps according to the relationships considering the foliage clumping and scale shift effect [13], and finally upscaled the 30 m LAI map to match with remote sensing LAI products [30]. Thus we have 53 field measurements in total, which include seven measurements in China. The distribution of global field measurement sites used for direct LAI validation is showed in Figure 3.

**Figure 3.** Distribution of field measurement LAI sites that are capable of direct LAI validation during 2001–2011.

## *2.8. Comparison Method*

The four products needed to be compared over the same spatial area and temporal period. In our study, we chose China as the study area because of its complex topography, various climate conditions and biome types; the three new LAI products were developed and maintained by Chinese groups and then expected to have reliable performance over China.

As stated in earlier research [22,24,29], direct validation is necessary to evaluate the accuracy of each product. Because the field measurements in China are too limited (only seven sites are available), we first validate the four products at global scale (53 measurement sits available), and then validate in China (seven measurement sites available). The results are shown in Section 3.1.

For LAI comparison, firstly, the four products were projected to the Albers projection coordinate system and resampled to 1 km by the nearest-neighbor sampling method for all the images during the overlapped period from 2001 to 2011. Then we calculated the yearly temporal mean LAI pixel by pixel for each year of each product as the basic data for comparison. The climatological LAI of each product was calculated by averaging the yearly temporal mean LAI from 2001 to 2011 and their spatial distribution was compared (Section 3.2.1). The difference among the four climatological LAI were presented in Section 3.2.2, while their spatial similarity was illustrated through scatter plots in Section 3.2.3. The standard deviation (SD) and relative standard deviation (RSD) were also analyzed in Section 3.2.4. SD and RSD are calculated as:

$$\text{SD} = \sqrt{\frac{1}{N}} \sum\_{i=1}^{N} \left( LAI\_i - \overline{LAI} \right)^2 \tag{1}$$

$$\text{RSD} = \frac{SD}{\overline{LAI}} \tag{2}$$

where *LAIi* represents the four LAI product value, LAI is the mean of four products, and *N* is the total number of products, in our study is 4.

Later, we computed the Pearson correlation coefficients (R) of the time series of yearly LAI pixel by pixel among four products and showed their spatial patterns in Section 3.3. R is computed as:

$$\mathcal{R} = \frac{n\sum\_{i=1}^{n} x\_i y\_i - \sum\_{i=1}^{n} x\_i \sum\_{i=1}^{n} y\_i}{\sqrt{n\sum\_{i=1}^{n} x\_i^2 - \left(\sum\_{i=1}^{n} x\_i\right)^2} \cdot \sqrt{n\sum\_{i=1}^{n} y\_i^2 - \left(\sum\_{i=1}^{n} y\_i\right)^2}}\tag{3}$$

where *xi*, *yi* represent two time series LAI value respectively in this case, and *n* represents the total number of year, which here is 11.

At last, we made the LAI difference case analysis, which included the mean SD/RSD and the LAI value difference for each biome type (Section 3.4.1), the proportion of each biome type at different SD/RSD significant levels and typical region case studies (Sections 3.4.2 and 3.4.3).

#### **3. Results**

## *3.1. Direct Validation with Field Measurement LAI*

We first validated the four LAI products: GLASS, GLOBALBNU, GLOBMAP, and MODIS to field measurement LAI. Considering the OLIVE sites are located in almost homogeneous land and the six China field measurements have been upscaled from high-resolution LAI images [30], we extracted the pixel value to directly match the field measurements. Furthermore, we chose images with the closest date to the ground measurement date for validation. Finally, we obtained 53 pixel LAI values for each LAI product for validation, and the uncertainty of each product was quantified by *R*2, *p*-value, and RMSE. Table 3 summarizes the validation indictors of the four products for global with 53 sites and China with seven sites, respectively. For global validation, GLASS shows the highest accuracy (*R*<sup>2</sup> = 0.70, RMSE = 0.96). For direct validation over China, the lowest uncertainty was achieved by GLASS LAI (*R*<sup>2</sup> = 0.94, RMSE = 0.61), while the highest uncertainty was obtained by MODIS LAI (*R*<sup>2</sup> = 0.03, RMSE = 2.12).


**Table 3.** Validation indictors of four products.

Taking biome types into consideration, different LAI products have different performance. Figure 4 shows the validation scatter plots over global 53 sites. For grassland (GRA), shrubland (SHR) and cropland (CRO), GLASS, GLOBALBNU and MODIS LAI all perform well and the scatter plots mostly stand on the 1:1 line. For evergreen needleleaf forest (ENF), the four products all show two sites of them perform better, while the other two sites perform worse. For evergreen broadleaf forest (EBF), GLASS and MODIS perform best, followed by GLOBALBNU and GLOBMAP. For mixed forest (MF), the four products are all overestimated (green triangles in Figure 4).

Some possible uncertainty sources could attribute to the difference between remote sensing LAI and field-measured LAI. First is the inversion errors that resulted from difference between remote sensing observation reflectance and modeled reflectance. The observation reflectance could be affected by several factors such as aerosol, cloud contamination, topography, etc., while modeled reflectance could be affected by the calibration parameters, and they both accumulate errors during inversion processes. In addition, a vegetation reflectance saturation problem could affect inversion accuracy, that is, reflectance is insensitive to dense canopies LAI [18], and the large LAI uncertainty of evergreen needleleaf forest in Figure 4 (red dot) could be attributed to this.

Second is the field-measured errors. The field LAI measurements are obtained by LAI-2000, TRAC, etc. instruments, but they do not distinguish the effective photosynthetic tissue from non-photosynthetic tissues such as branches, stalks, and dead leaves, which could overestimate the LAI value [34]. The results in Figure 4 that remote sensing LAI is lower than field-measured LAI could be due to this reason. For some dense biomes, the field-measured LAI have not considered the underforest canopy, while remote sensing LAI is the observation of the vegetation vertical structure, which considers the underforest [35]. The overestimate of mixed forest in Figure 4 may result from this.

**Figure 4.** Directly validated scatter plots: (**a**) GLASS; (**b**) GLOBALBNU; (**c**) GLOBMAP; (**d**) MODIS. MF represents mixed forest, ENF represents evergreen needleleaf forest, EBF represents evergreen broadleaf forest, SHR represents shrubland, GRA represents grassland, CRO represents cropland.

Spatially, although most sites we chose are at homogenous land, but there are still some mixed land cover types and the scale effect is inevitable. Temporally, the TM/ETM+ images and field sampling date may be different during the generation process of fine-resolution LAI maps. In addition, the remote sensing LAI images are for dates closest to the field-measured dates. These are another two error sources that lead to validation uncertainty.
