Vegetation variable retrieval from reflectance data is typically grouped into three categories: the statistical-empirical category, the physical category and the hybrid category (physical models applied to statistical models). Based on the similarities between the spectra of leaves in the optical domain, the leaf reflectance spectra can be linearly modelled using a very limited number of principal components (PCs) if the PCA (principal component analysis) transformation is carried out at the sample dimension. In this paper, we present a novel data-driven approach that uses the PCA transformation to reconstruct leaf reflectance spectra and also to retrieve leaf biochemical contents. First, the PCA transformation was carried out on a training dataset simulated by the PROSPECT-5 model. The results showed that the leaf reflectance spectra can be accurately reconstructed using only a few leading PCs, as the ten leading PCs contained 99.999% of the total information in the 3636 training samples. The spectral error between the simulated or measured reflectance and the reconstructed spectra was also investigated using the simulated and measured datasets (ANGERS and LOPEX'93). The mean root mean squared error (RMSE) values varied from 5.56 × 10 −5 to 6.18 × 10 −3 , which is about 3-10 times more accurate than the PROSPECT simulation method for measured datasets. Secondly, the relationship between PCs and leaf biochemical components was investigated, and we found that the PCs are closely related to the leaf biochemical components and to the reflectance spectra. Only when the weighting coefficient of the most sensitive PC was employed to retrieve the leaf biochemical contents, the coefficients of determination for the PCA data-driven model were 0.69, 0.99, 0.94 and 0.68 for the specific leaf weight (SLW), equivalent water thickness (EWT), chlorophyll content (Cab) and carotenoid content (Car), respectively. Finally, statistical models for the retrieval of leaf biochemical contents were developed based on the weighting coefficients of the sensitive PCs, and the PCA data-driven models were validated and compared to the traditional VI-based and physically-based approaches for the retrieval of leaf properties. The results show that the PCA method shows similar or better performance in the estimation of leaf biochemical contents. Therefore, the PCA method provides a new and accurate data-driven method for reconstructing leaf reflectance spectra and also for retrieving leaf biochemical contents.

The fraction of absorbed photosynthetically active radiation by vegetation (FAPAR) is a key variable in describing the light absorption ability of the vegetation canopy. Most global FAPAR products, such as MCD15A2H and GEOV1, correspond to FAPAR under black-sky conditions at the satellite overpass time only. In this paper, we aim to produce both the global white-sky and black-sky FAPAR products based on the moderate resolution imaging spectroradiometer (MODIS) visible (VIS) albedo, leaf area index (LAI), and clumping index (CI) products. Firstly, a non-linear spectral mixture model (NSM) was designed to retrieve the soil visible (VIS) albedo. The global soil VIS albedo and its dynamics were successfully mapped at a resolution of 500 m using the MCD43A3 VIS albedo product and the MCD15A2H LAI product. Secondly, a method based on the energy balance residual (EBR) principle was presented to retrieve the white-sky and black-sky FAPAR using the MODIS broadband VIS albedo (white-sky and black-sky) product (MCD43A3), the LAI product (MCD15A2H) and CI products. Finally, the two EBR FAPAR products were compared with the MCD15A2H and Geoland2/BioPar version 1 (GEOV1) black-sky FAPAR products. A comparison of the results indicates that these FAPAR products show similar spatial and seasonal patterns. Direct validation using FAPAR observations from the Validation of Land European Remote sensing Instrument (VALERI) project demonstrates that the EBR black-sky FAPAR product was more accurate and had a lower bias (R2 = 0.917, RMSE = 0.088, and bias = −2.8 %) than MCD15A2H (R2 = 0.901, RMSE = 0.096, and bias = 7.6 % ) and GEOV1 (R2 = 0.868, RMSE = 0.105, and bias = 6.1%).