*3.3. Temporal Correlation of Time Series of Annual Mean LAI*

Figure 11 illustrates the temporal correlation among six pairs of annual mean LAI. For this 11-year time series, the correlation coefficients equal to ±0.48, ±0.55, and ±0.68 correspond to a 10%, 5%, and 1% significance level, respectively. For the six pairs, the significant correlation regions are mainly located in southern China (red color), which means the four products have high consistency in these regions. However, in most regions of southern China, we could not find a significant correlation (green). Specifically, GLOBALBNU and MODIS (Figure 11b) have a significant correlation in almost the whole of China, which reflects the temporal similarity of these two products.

**Figure 11.** Temporal correlation among each pair of products: (**a**) GLASS vs. MODIS; (**b**) GLOBALBNU vs. MODIS; (**c**) GLOBMAP vs. MODIS; (**d**) GLASSGLOBALBNU; (**e**) GLASS vs. GLOBMAP; (**f**) GLOBALBNU vs. GLOBMAP.

 vs.

From the above analysis, we know the four LAI products have their own characteristics and differences. Although GLASS, GLOBALBNU, GLOBMAP, and MODIS LAI products are all inversed from MODIS land surface reflectance product MOD09, differences still exist among these four LAI products, which can be caused by several key factors. First is the process of input land surface reflectance. Although the product has been atmospherically corrected, there still exists some residual contamination such as aerosol, cloud, snow, etc. In view of this, GLASS LAI has produced a 500-m cloud and snow mask to reprocess the MOD09 product according to a method proposed by Tang [36]. GLOBALBNU LAI maintains the pixels of MODIS QC = 0, which means pixels with no cloud and snow, and makes a two-step spatial and temporal filter for pixels QC > 0. GLOBMAP was filtered by MOD09 cloud mask layer, and MODIS adopted a back-up algorithm when the main algorithm failed because of cloud contamination.

The second factor is the LAI retrieval algorithm. There are two main methods to retrieve LAI. One is empirical and is based on the relationship between LAI and VI. The other is a physical method that relies on the reversion of a canopy radiative transfer model, such as the MODIS look-up-table method and artificial neural networks [29]. GLASS LAI uses general regression neural networks to generate a yearly LAI product from time series MODIS reflectance, which belongs to a physical method. GLOBMAP integrates MODIS reflectance and the GLOBCARON algorithm to retrieve improved LAI from 2000 to 2011. To retrieve LAI back to 1981, a pixel-by-pixel relationship is established between improved LAI and AVHRR NDVI, which is an empirical method. MODIS LAI employs look-up tables simulated from a 3D radiative transfer model, and GLOBALBNU LAI filters the MODIS LAI. These are both physical methods.

The third factor can be attributed to land cover maps. In these four LAI retrieval systems, four different land cover inputs are used. GLASS classifies the biome into eight types according to the MCD12Q1 type 3 layer (grass and cereal crop, shrub, broadleaf crop, savanna, evergreen broadleaf, deciduous broadleaf, evergreen needleleaf, deciduous needleleaf). GLOBALBNU LAI uses the MCD12Q1 type 5 layer, without cereal crop and savanna but with increased barren and sparse vegetation type. GLOBMAP LAI uses the MCD12Q1 type 1 layer, which classifies the biome into six types, including grass and cereal crop, conifer forest, tropical forest, deciduous forest, mixed forest, and shrub, and MODIS classifies biome into grass and cereal crop, shrub, broadleaf crop, savanna, broadleaf forest, and needleleaf forest, and separates cereal crops and broadleaf crops. Myneni et al. [12] estimated that classification errors in land cover maps can generate an LAI estimation error of up to 50%, thus land cover types in an LAI retrieval system play an important role and should not be neglected.

Other factors can be related to a clumping index. For field measurement of LAI data, different measurement methods can lead to different LAI. For instance, LAI derived by direct and destructive measurement can be considered true LAI, while indirect methods such as LAI2000 instrument can be regarded as effective LAI. The distinction between true LAI and effective LAI is whether the measurements take the foliage's spatial distribution into account, for we assume the plant canopy architecture is under random distribution. The clumping index is the measure of foliage grouping relative to a random distribution of leaves in space [37], which provides the conversion between effective and true LAI. However, the four LAI products used in our study adopt different clumping indexes. GLASS LAI uses a clumping index map derived from POLDER 3 [38], which has removed the topographical effect that leads to a cross-biome difference [24], while GLOBALBNU and GLOBMAP LAI uses the POLDER 1 clumping index developed by Chen [39].
