Generation of a 16 m/10-day fractional vegetation cover product over China based on Chinese GaoFen-1 observations: method and validation

ABSTRACT

As China has recently launched the GaoFen-1 satellite (GF-1) carrying on the wide-field view (WFV) sensor, it is a challenging task to make full use of its observations to produce the fractional vegetation cover (FVC). In light of this, our study presents a comprehensive algorithm to generate a 16 m/10-day FVC product by considering the vegetation types characteristics. For forests, considering the foliage clumping effect, FVC was estimated from the gap probability theory using GF-1 leaf area index (LAI) and clumping index (CI) as a priori knowledge; for non-forests, FVC was estimated from the dimidiate pixel model using GF-1 normalized difference vegetation index (NDVI). The performance of GF-1 FVC from 2018 to 2020 was evaluated using FVC ground measurements obtained from 7 sites for crops, grasslands, and forests in China. The direct validation indicated that the performance of the FVC product was satisfactory, as evidenced by R²= 0.55, RMSE = 0.15 and BIAS = 0.01 for all vegetation types. Furthermore, the GF-1 FVC exhibited better performance compared to the GEOV3 FVC at a spatial resolution of 300 meters. Moreover, the 10-day temporal interval of GF-1 FVC product successfully facilitated the extraction of regional phenological information at a spatial resolution of 16 meters.

KEYWORDS:

• FVC
• GF-1 WFV
• gap probability theory
• dimidiate pixel model

1. Introduction

Vegetation is widely acknowledged as a fundamental element within terrestrial ecosystem, thus vegetation dynamics is a key factor in monitoring ecosystems (Bian et al. 2017). Fractional vegetation cover (FVC), defined as the fraction of the landscape occupied by the vertical projection of green vegetation per unit area (Gitelson et al. 2002), is a critical parameter for monitoring vegetation growth. Moreover, FVC serves as a vital indicator of the ecosystem, facilitating the interactions between terrestrial and atmospheric ecosystems, as well as their hydrological processes (Jiapaer, Chen, and Bao 2011).

Previous studies have demonstrated the successful utilization of optical remote sensing data for the mapping of long-term FVC mapping at large scale (Delamater et al. 2012; Guo et al. 2015). The spatial resolution of most existing FVC products at the global scale is coarser than 1 km, including POLDER, CYCLOPES, GEOV, ENVISAT and Multi-source data Synergized Quantitative remote sensing production system (MuSyQ) FVC products (Bacour et al. 2006; Baret et al. 2007; Jia et al. 2015; Mu et al. 2017). The Global LAnd Surface Satellite (GLASS) FVC product has a spatial resolution of 500 m (Jia et al. 2015; Jia et al. 2019), while the global GEOV3 FVC from Copernicus Global Land Service (CGLS) have a spatial resolution of 300 m (Fuster et al. 2020). However, climate change analysis and regional scale hydrological modeling require FVC products with spatial resolution higher than 100 m (Yang et al. 2017), which highlights the need to generate new FVC products with decametric resolution.

Since 1984, a range of satellite sensors with medium- to high-resolution have become accessible. The Landsat series sensors (including thematic mapper (TM)/enhanced thematic mapper plus (ETM+)/operational land imager (OLI)), SPOT high resolution geometric (HRG), Sentinel-2 and PlanetScope sensors have been used to extract FVC at regional scales (Andreatta et al. 2022; Ding et al. 2016; Dubovyk et al. 2015; Riihimaki, Luoto, and Heiskanen 2019; Yang et al. 2017; Zhang, Harris, and Balzter 2015; Zhou, Yang, and Chen 2018; Zribi et al. 2016). Currently, two regional FVC products are available at decametric resolution and spanning more than one year, including a 30 m/15d FVC product over China in 2010 derived from Chinese HJ-1 data and MODIS data (Mu et al. 2021), and a 30 m/16d FVC product over Hubei Province in 2017 derived from Landsat and Sentinel-2 data (Wang et al. 2023). However, the sensitivity of cloud contamination, limited scan width (290 km for Sentinel-2, 185 km for Landsat, 60 km for both ASTER and SPOT) and long revisiting period (10 days for Sentinel-2, 16 days for ASTER and Landsat, and 26 days for SPOT) of the medium- to high-resolution satellite observations make it difficult to obtain high-quality continuous spatiotemporal estimates at regional scales. In 2013, China launched the 16 m resolution GaoFen-1 satellite (GF-1) with a wide-field view (WFV) camera. The combined GF-1 images from 4 WFV cameras have a scan width of about 800 km over a 4-day revisiting time; in contrast to the Landsat satellites with a scan width of 185 km over a 16-day revisiting time. The GF-1 WFV data have the potential to provide more frequencies and higher quality observations for FVC estimates. Previous researches have proved that FVC estimates from GF-1 WFV show reasonable patterns at the regional scale (Jia et al. 2016; Li et al. 2016; Song et al. 2022; Tao et al. 2019; Zhan et al. 2014).

Methods for FVC estimation from medium- to high-resolution images include the spectral mixture analysis (SMA) method (Li et al. 2016; Zhang, Harris, and Balzter 2015), the machine learning algorithm (i.e. deep learning approach, backward propagation neural networks (BPNNs), multivariate adaptive regression splines (MARS)) based on the radiative transfer model (PROSAIL, canopy reflectance model for row crops, etc.) (Jia et al. 2016; Ma et al. 2021; Wang et al. 2023; Yang et al. 2017), and the dimidiate pixel model method based on normalized difference vegetation index (NDVI) or new vegetation index from red-edge band (Deng et al. 2021; Mu et al. 2021; Song et al. 2022; Wang et al. 2017) from Landsat TM/ETM+/OLI, Sentinel-2, World-View 3, HJ-1 CCD, GF-1/6 WFV, and ZY-3. Among these methods, the end members of the SMA method and the training samples of the neural network (NN) method rely on sample selection, which requires calibration of the coefficients when applied to large scales. The highly simplified formulation in the dimidiate pixel model is available to estimate FVC with medium- to high-resolution images at regional scales (Gao et al. 2020; Zhou, Yang, and Chen 2018). The NDVI, which is sensitive to vegetation growth, is commonly used in the dimidiate pixel model method. The only concern is that the NDVI values easily reach saturation for dense canopy (Liu, Qin, and Zhan 2012). Moreover, the use of a new red-edge index with higher sensitivity to FVC is limited by the availability of the red-edge band in satellite sensors. Currently, only the Sentinel-2 and GF-6 satellites include the red-edge band, which limits the generation of long-term FVC products. Additionally, due to the diverse characteristics of tree clusters within forests, including varying height, densities, and crown gaps, the forest canopy in medium- to high-resolution imagery frequently demonstrates notable spatial discreteness, in contrast to the relatively uniform and homogeneous canopy observed in agricultural fields and and grasslands. The dimidiate pixel model is no longer applicable for characterizing vegetation with dense and complex structures, such as forests. Instead, the utilization of the gap probability theory, which incorporates the effects of canopy clumping, presents an alternative approach for estimating FVC for forests at large scales (Zhao et al. 2020). However, the leaf area index (LAI) with a smaller value for the leaf expansion stage will lead to an underestimation of FVC from the gap probability theory for crop and grass types (Carlson and Ripley 1997; Xiao et al. 2016). Therefore, it is necessary to select an appropriate estimation method with canopy structure for different vegetation types (Yin et al. 2015).

The objective of this study is to present a practical approach for producing FVC data for China using GF-1 WFV data with a spatial resolution of 16 m and a temporal resolution of 10 days, covering the period from 2018 to 2020. Considering the distinct characteristics of forests and non-forests, the gap probability theory was employed to estimate FVC for forests, utilizing GF-1 LAI product and clumping index (CI) as a priori-knowledge. Additionally, and the dimidiate pixel model was utilized to estimate FVC for non-forests, based on GF-1 NDVI product. The analysis focused on examining the spatial and temporal consistency of GF-1 FVC estimates. Furthermore, FVC estimates derived from ground FVC measurements across China were evaluated and compared to the GEOV3 FVC product.

2. Materials

2.1. Remote sensing products

2.1.1. GF-1 NDVI and LAI products

The GF-1 WFV satellite data were provided by the China Centre for Resources Satellite Data and Application (CRESDA) (http://www.cresda.com). The top of canopy reflectance was generated using the Analysis Ready Data (ARD) method (Zhong et al. 2021). The NDVI was then calculated from the red band and near-infrared band reflectance on each observation date and the final GF-1 NDVI product was derived by selecting the maximum NDVI during each 10 days to further reduce the effect of cloud contamination. The LAI was first retrieved from a biome-specific look-up table (LUT) based on the three-dimensional radiative transfer model (3D-RTM) for each GF-1 WFV image (Xu et al. 2020), and then the maximum LAI with high quality within the 10 days was selected as the GF-1 LAI product value. Both 16 m/10-day GF-1 LAI and NDVI products were produced by the MuSyQ system with the Aerospace Information Research Institute, Chinese Academy of Sciences (Li et al. 2021; Zhang et al. 2021).

2.1.2. Land-cover map

The fine land-cover map with 30 m spatial resolution in 2015 (Zhang et al. 2019), acquired from the National Earth System Science Data Center, National Science and Technology Infrastructure of China (http://www.geodata.cn), was employed to identify the vegetation types before FVC estimation. The vegetation classifications include cropland, broad-leaf forests, coniferous forests, mixed forests, shrubland, grassland, and wetlands. The classifications of impervious, bare land, water bodies, snow, and ice were excluded from in the subsequent processing.

2.1.3. Climate classification map

The global climate classification of Köppen-Geiger system includes five main classes (i.e. tropical (A), arid (B), temperate (C), cold (D), and polar (E) climates) and more than 30 subtypes. To derive a priori-knowledge of CI and NDVI coefficients for various vegetation types, the global Köppen-Geiger climate classification map was employed, which was generated by considering thresholds and seasonality of monthly air temperature and precipitation (Beck et al. 2018; Kottek et al. 2006).

2.1.4. CI product

A CI product with a spatial resolution of 500 m and a temporal resolution of 8 days, estimated based on the MODIS BRDF product using a linear relationship between CI and the normalized difference between hotspot and darkspot (NDHD) angular index (Jiao et al. 2018), was used to extract CI priori knowledge for broad-leaf forests, coniferous forests, mixed forests and shrubs under five major Köppen-Geiger system climate types.

2.1.5. GEOV3 FVC product

The GEOV3 FVC product is an extension of the standard GEO version 1 and 2 FVC products, which were derived from SPOT-VEGETATION sensors. The GEOV3 FVC product is generated using the Sentinel-3/OLCI or PROBA-V sensors, with a temporal interval of 10 days and a spatial resolution of 300 m on a global scale (Fuster et al. 2020). The generation of the GEOV3 FVC product involves the use of a neural network that combines simulations from the PROSPECT leaf optical properties model with the Scattering by Arbitrarily Inclined Leaves (SAIL) canopy bidirectional reflectance model, abbreviated as PROSAIL. To ensure the continuity and consistency of the FVC product, the gap-filling and temporal smoothing method was implemented on the generated FVC estimates based on the climatology information (Dierckx et al. 2014).

2.2. Ground sites for evaluation

2.2.1. Ground measurements in China

A total of 579 ground measurements (including 110 for crops, 75 for forest and 394 for grass) were available throughout the growing season in China to derive a comprehensive evaluation of the GF-1 FVC product from the Huailai site, Hebei Province, in 2014; the middle reaches of the Heihe River Basin, Gansu Province, the Jingyuetan site, Jilin Province and the Saihanba Plantation Forestry Farm, Hebei Province, in 2018; the study area of Beibei, Chongqing City and Pu'er, Yunnan Province in 2020; and the Hulunbuir Steppe, Inner Mongolia Autonomous Region, in 2021. The spatial distribution of ground measurements is shown in Figure 1. Detailed information for each field experiment, including the geographic location, dominant species, observation dates, and the number of FVC ground measurements, are reported in Table 1.

Figure 1. Geographic locations of ground measurement sites and the comparison sites across China are depicted in the dot samples. These samples represent the ground measurements of FVC over the GF-1 false-color composite image, utilizing near-infrared, red, and green bands for each study area in subgraphs. The yellow dots indicate crop areas, and the sugilite sky dots indicate forest areas. The yellow dots over a light green background represent grass type samples. The colorful star and triangle samples denote the ground measurement sites used for inter-comparisons.

Table 1. Details of the FVC ground measurements used for evaluation.

Site names/ field areas	Latitude/Longitude	Dominant species	Observation dates	Number of plots
Huailai	40°20′55.09′′N, 115°46′59.57′′E	Crops (Maize)	2014.05-07	4
Heihe	39°35′59″N, 99°24′8″E	Crops (Maize)	2018.05-09	30
Jingyuetan	43°47′24.55′′N, 125°26′ 41.45′′E	Crops (Maize/soybean)	2018.07-12	30
Saihanba	42°24′5.72′′N, 117°16′49.84′′E	Forest (Pine/white birch)	2018.05-08	11
Beibei	29°44′58′′N-29°47′26′′N, 106°18′47′′E-106°26′52′′E	Crops/forest (fruit and tea forest)	2020.10-11	37/41
Pu’er	22°44′4′′N-22°51′24′′N, 100°55′43′′E-100°56′57′′E	Crops/forest (tea and natural forest)	2020.12	9/23
Hulunbuir	47°05′N-53°20′N, 115°31′−126°04′E	Grass	2021.07-08	394

The in-situ data were observed from the digital camera photographs for Huailai station, the middle reach of the Heihe River basin, Saihanba plantation forestry farm, Beibei, and Pu'er field experimental areas. The sample size is 10m × 10 m or 20m × 20 m for each station. The in-situ data were collected by PhenoCam for Jingyuetan station at five fixed points. All digital images obtained from both the digital camera photographs and PhenoCam were processed by an algorithm based on a Gaussian distribution for both vegetated and non-vegetated parts, taking into account the viewing angle of the digital image (Liu et al. 2012a; Song et al. 2015). The mean error associated with the FVC extracted by this method and the supervised classification method was less than 0.035 (Liu et al. 2012aa), which gives us confidence in the good reliability of field FVC measurements. In-situ data on grass type in the Hulunbuir steppe region were obtained by an acupuncture method, which measured the proportion of grass in a 1m × 1 m quadrat using a probe.

2.2.2. Inter-comparison site for GF-1 and GEOV3 FVC products in China

A total of 35 homogeneous ground experiment sites were provided by CERN (Chinese Ecosystem Research Network) and BELMANIP-2.1 (BEnchmark Land Multisite ANalysis and Intercomparison of Products) (Baret et al. 2006) in China. These sites included 12 forest types and 23 non-forest types (e.g. 20 crops and 3 grass types). The 24 sites from BELMANIP-2.1 (Figure 1 colored triangle) were selected with a homogeneous area of 10km × 10 km. In addition, the 11 sites from CERN (Figure 1 colored stars) were selected with a homogeneous surface over a 1km × 1 km area based on high spatial resolution imagery (Xu et al. 2016).

3. Methodology

3.1. Method of GF-1 FVC estimation

3.1.1. Dimidiate pixel model method

The formula of the dimidiate pixel model can be expressed as: (1) $FVC={\displaystyle \frac{NDVI-NDV{I}_{min}}{NDV{\mathrm{I}}_{max}-NDV{\mathrm{I}}_{min}}}$(1)

Two key parameters in the dimidiate pixel model (i.e. the NDVI values for high density vegetation (NDVI_veg) and for bare soil (NDVI_soil)) determined the accuracy of the FVC estimates. Considering the variations of NDVI_veg and NDVI_soil in different climate types and plant growth conditions (Gao et al. 2020; Montandon and Small 2008), this study extracted the maximum and minimum values of 16 m/10-day GF-1 NDVI from the time series for different vegetation types and climate zones. These extracted values were then used to generate the LUT coefficients for NDVI_veg and NDVI_soil.

3.1.2. Gap probability theory method

The feasibility of FVC generation based on gap probability theory was demonstrated in our previous study (Zhao et al. 2020) based on Terra and Aqua MODIS images with a spatial resolution of 500 m every 8 days in Europe. The formula of the gap probability theory method is as follows: (2) $FVC(\theta )=1-\exp (-LAI\cdot \mathrm{\Omega }\cdot G/\cos (\theta ))$(2)

The input parameters, i.e. LAI using the GF-1 LAI product, the view zenith angle (θ), which is closed to the nadir observation (0°) from the high spatial resolution image, was set to 0, and the G-function coefficient was set to 0.5 in this study. The Ω describes the foliage clumping effect. However, there is currently no available CI product with a 16 m spatial resolution that can be directly used in this method. Therefore, in this study, a priori-knowledge of CI values for different vegetation types was parameterized. For this purpose, homogeneous pixels were defined as those with the same vegetation type within a spatial resolution of 500 m, using information from the land-cover map. In addition, considering the spatial clumping within the vegetation and canopy, which varies with different temperature and precipitation conditions, a CI priori knowledge LUT was generated as the mean value within each 8-day CI product for different vegetation types, considering homogeneous pixels and different climate zones, during the period from 2018 to 2020.

3.1.3. Implement

The workflow for generating the 16 m/10-day FVC product is shown in Figure 2. Based on the land-cover and climate classification maps, we implemented the gap probability theory method for forest types (i.e. broad-leaf forests, coniferous forests, mixed forests and shrubs) based on the 16 m/10-day GF-1 LAI and CI priori-knowledge, and implemented the dimidiate pixel model method for non-forest types (i.e. cropland, grassland and wetlands) based on the 16 m/10-day GF-1 NDVI. Once each inversion method has been successfully run, the FVC estimates for both forest and non-forest types were recorded, otherwise, the estimate was filled with an invalid value. Quality flags were also recorded along with the FVC estimates. Based on the above method, a dataset of 16 m/10-day FVC product from 2018 to 2020 over China was released on the Science Data Bank website. The details of the dataset profile and data preprocessing can be found in Zhao et al. (2022).

Figure 2. Flowchart of generating a 16 m/10-day FVC product from GF-1 satellite data.

3.2. Assessment and evaluation

3.2.1. Direct evaluation

The accuracy of the 16 m/10-day GF-1 FVC product was evaluated by the ground measurements with three vegetation types over China (Figure 1 and Table 1). To increase the sample numbers for evaluation, FVC estimations from periods outside of 2018–2020 in Table 1 were also included in this study. The FVC estimate closest in date to the corresponding ground measurements in Table 1 was selected for each site. The performance of the FVC estimates based on the ground measurements was assessed using three indices: R², RMSE, and BIAS.

3.2.2. Assessment analysis

3.2.2.1. Spatial and temporal continuity

The spatial and temporal continuity of the 16 m/10-day FVC product is demonstrated by analyzing the fraction of missing data in both space and time, considering factors such as cloud or shadow contamination, inadequate data preprocessing outcomes, and other technical challenges encountered during image acquisition. The missing fraction index represents the proportion of invalid estimates within each synthesis period throughout the entire year. All the data available in 2019 were utilized at a spatial resolution of 16 m and a temporal resolution of 10-day. To conduct the spatial continuity analysis, the spatial distribution of missing ratios was assessed across China in 2019. Subsequently, the proportion of missing data was examined about both time and latitude. Regarding the temporal continuity analysis, the fluctuation patterns of the missing product rate were explored across various seasons as a function of vegetation type.

3.2.2.2. Spatial and temporal consistency

To ensure the comparability of GF-1 and GEOV3 FVC products, the FVC estimates derived from GF-1 WFV were initially aggregated to a spatial resolution of 300 m. Consequently, the temporal consistency between GF-1 FVC and GEOV3 FVC was assessed using identical temporal (10 days) and spatial (300 m) resolutions. The seasonal variability of forest, crop, and grass types at ground measurement sites was examined by analyzing the temporal profiles of GF-1 FVC and GEOV3 FVC products.

Furthermore, this study incorporates an assessment of the two FVC products through quantitative analysis using ground FVC measurements and conducting a comparative analysis with inter-comparison sites featuring homogeneous surfaces. The performance of the FVC products is evaluated based on metrics such as R², RMSE and bias.

Considering that the multi-temporal ground measurements were obtained at the sample size of 10m × 10 m or 20m × 20 m, some ground measurements would be located in the same pixels at 300 m spatial resolution. The mean FVC of the ground measurements within 300 m spatial resolution for homogeneous pixel based on Sentinel images was used as the evaluation truth value. At the same time, the processed verification truth values and two FVC products with valid FVC values were selected for evaluation. After processing, a total of 418 verification truth values (including 35 samples for crops, 38 samples for forests and 345 samples for grass) from multi-temporal ground measurements at 300 m spatial resolution were included in the accuracy evaluation.

For inter-comparison, the aggregated FVC of the GF-1 product and the GEOV3 FVC product were both averaged over one month to eliminate the difference in temporal composites. Considering the ground measurements with multi-temporal observation dates, about 401 ground measurements with a single geographical location and observation date and 35 homogeneous ground-based experimental sites from Chinese CERN and BELMANIP-2.1 were used to implement the inter-comparison between these two FVC products from 2018 to 2020.

4. Results

Figure 3 depicts the spatial patterns of GF-1 FVC across China for different days of the year (DOY) in 2019. The observed pattern of FVC aligns with the existing vegetation types in China. From the January (DOY 001/011/021) to July (DOY 181/191/201), there is a gradual rise in FVC from the northwest to the southeast. The northeastern and southeastern regions of China exhibit high vegetation density, whereas the northwestern region of China showcases extensive desert areas with FVC values approaching zero. Furthermore, Figure 4 illustrates regional maps comparing GF-1 FVC at a spatial resolution of 16 m with GEOV3 FVC at a spatial resolution of 300 m for both forested and cultivated areas during the same period. It's worth noting that GF-1 FVC provides more detailed information compared to GEOV3 FVC.

Figure 3. Spatial distribution map of GF-1 FVC products across in China in 2019 for different DOYs. (a)-(c) for January, and (d)-(f) for July. (a) DOY 001; (b) DOY 011; (c) DOY 021; (a) DOY 181; (b) DOY 191; (c) DOY 201.

Figure 4. Comparison details of GF-1 FVC at 16 m spatial distribution compared with GEOV3 FVC at 300 m spatial distribution in 2019 for forest and crops types at regional areas. (a) DOY 011 for forest type; (b) DOY 191 for forest type; (c) DOY 011 for crops type; (d) DOY 191 for crops type.

4.1. Direct evaluation of GF-1 FVC with 16m spatial resolution

Figure 5 shows the direct evaluation result between the GF-1 FVC estimates and FVC ground measurements with a total of 579 ground truth measurements from 7 ground measurement sites over China. The accuracy of GF-1 FVC, with R² of 0.55, RMSE of 0.15 and BIAS of 0.01 for all vegetation types, showed reasonable results with FVC ground measurements (Figure 5). The grass type exhibited the most exceptional performance, with an RMSE value of 0.12, followed by the forest and crop types.

Figure 5. Scatterplot of FVC ground measurements and GF-1 FVC estimates at 7 ground measurement sites in China.

4.2. Spatial and temporal continuity of the GF-1 FVC product at 16m spatial resolution

4.2.1. Spatial continuity

The spatial distribution of the FVC missing ratio for 2019 shows a significant difference from north to south China (Figure 6). About 47.07% of mainland China had missing ratios below 30%, which were mainly distributed in northern China with cloudless and clear weather conditions (Figure 6, blue areas). Annual missing values between 40% and 70% accounted for 38.29% of the total mainland China, distributed in the north and northwest of the country. The higher percentage of missing data (≥80%) accounted for 14.64% of mainland China, distributed at lower latitudes in south-eastern China (Figure 6, yellow areas), mainly due to cloud contamination, rain effects throughout the year, or affected by poor data pre-processing, resulting in no valid data.

Figure 6. Spatial distribution of the missing ratio for GF-1 FVC over China in 2019.

4.2.2. Temporal continuity

Figure 7 shows the missing values of the GF-1 FVC products in different seasons, which vary with latitude in 2019. The quarters of this year (Q1–Q4) were divided into four periods from January to December. The percentage of missing values in the middle to high latitudes from 40.3° to 53.4° is lower than ≤40%, which is strongly related to the better weather conditions with more cloud-free transit images from GF-1. However, the high percentage of missing values (>70%) from 18.2° to 30° is mainly caused by the heavy cloudiness and rain throughout the year, reducing the number of cloud-free transits from GF-1. The percentage of missing values varies from season to season. The missing values at mid-latitudes (22°∼43°) do not show a significant seasonal variation. In general, the range between the maximum and minimum missing ratio for different latitudes is larger for all four quarters. Moreover, the missing ratio is significantly related to the climatic characteristics in which it is located, i.e. the difference between the missing ratios at mid to high latitudes (43°∼53°) is more significant in Q1 and Q4 than that of Q2 and Q3, and the missing values at low latitudes (18°∼22°) are lower in Q1 than others.

Figure 7. Percentage of missing values of the GF-1 FVC product over land pixels in China as a function of latitude in 2019. The mean missing percentages for each quarter are shown as colored lines, and the ranges between the maximum and minimum missing percentages are shown as colored areas.

Figure 8 shows the average missing values for GF-1 FVC with different vegetation types (i.e. broad-leaf forests, coniferous forests, mixed forests, shrubs, crops and grass) in each season over China. The FVC estimated by the dimidiate pixel model based on GF-1 NDVI for crop and grass types had a low missing ratio (<10.6%) for four seasons, while the FVC estimated by the gap probability theory method based on GF-1 LAI for broad-leaf forests, coniferous forests and shrubs types had a higher missing ratio, ranging from 20% to 32%. Note that the highest missing ratio occurs for mixed forests, where the missing ratio is greater than 33% for all seasons. With the exception of crops and grass types, the average missing ratio for other vegetation types was higher in the third quarter (i.e. summer) than in other quarters.

Figure 8. Percentage of missing values for GF-1 FVC products with different vegetation types over China in 2019.

4.3. Comparison analysis between aggregated GF-1 and GEOV3 FVC at 300m spatial resolution

4.3.1. Distribution of FVC difference between two products

To ensure the comparability between the two FVC products, the pixels with valid values from both aggregated GF-1 and GEOV3 FVC were selected to participate in the comparison. From the spatial distributions of FVC difference between aggregated GF-1 and GEOV3 over China in January and July, 2019 (Figure 9), the aggregated FVC values of GF-1 are higher than those of GEOV3 in January from the northeast to southwest areas, but lower for the south of China (Figure 9(a)). On the contrary, the aggregated FVC values of GF-1 are lower than those of GEOV3 in July from the longitude of 100°E (Figure 9(b)).

Figure 9. Spatial distribution of FVC difference between aggregated GF-1 and GEOV3 over China in January and July, 2019. (a) January; (b) July.

Figure 10 shows a box plot of the difference between the aggregated GF-1 and GEOV3 FVC for different vegetation types. The smallest differences were observed for the shrub type in both January and July, followed by the grass type. Overall, the values of aggregated GF-1 FVC are greater than those of GEOV3 for all vegetation types in January, but show a decreasing trend in July. The absolute values of median FVC for all vegetation types were lower in July than in January.

Figure 10. Box plots of the difference between aggregated GF-1 and GEOV3 FVC for different vegetation types over China in January and July 2019. The box extends from the 25th to the 75th percentile. The median is shown as a line, and the mean is shown as a square. The bars correspond to the interquartile range of 1.5. Outliers are not shown.

4.3.2. Temporal consistency between aggregated GF-1 and GEOV3 FVC

Figure 11 shows the temporal profiles of aggregated GF-1 and GEOV3 FVC products with the 300 m spatial resolution and 10-day interval from 2018 to 2020 for forest, crops and grass types. All temporal profiles of three typical vegetation types show seasonal variation, although the ranges of variation for the forest and crop types are much larger than for the grass type. Compared with ground measurements, the aggregated GF-1 FVC showed better agreement with ground measurements of maize than the GEOV3 FVC at Heihe station (Figure 11(b)). Except for a few missing values during the growing season, the temporal profiles of GF-1 FVC can still reflect the phenological information for different vegetation types at 16 m spatial resolution.

Figure 11. Temporal profiles of FVC products for typical vegetation types over three field experiment regions. (a) Forest type in Saihanba; (b) Corn type in Heihe; (c) Grass type in Hulunbuir.

4.3.3. Consistency between GF-1 and GEOV3 FVC products

Figure 12 shows the evaluation results between the aggregated GF-1 or GEOV3 FVC products and FVC ground measurements at 300 m spatial resolution, for a total of 418 evaluation ground truth measurements from 7 ground measurement sites over China. The performance of aggregated GF-1 FVC (R²= 0.57, RMSE = 0.12, BIAS = −0.03) with FVC ground measurements was better than GEOV3 FVC (R²= 0.15, RMSE = 0.25, BIAS = −0.18), especially for grass (Figure 12 green dots).

Figure 12. Scatter plot of FVC ground measurements and FVC from aggregated GF-1 and GEOV3 at 7 ground sites over China. (a) Aggregated GF-1 FVC; (b) GEOV3 FVC.

Figure 13 shows a total of 35 homogeneous sites from BELMANIP-2.1 and CERN and 401 ground sampling points used to extract the aggregated GF-1 and GEOV3 FVC at 300 m resolution per month from 2018 to 2020. The FVC values show a scatter distribution for forest type, with the GEOV3 FVC having a higher value than the aggregated GF-1 FVC at a density of 0.8 (Figure 13(a)). The aggregated GF-1 FVC shows a slightly higher trend than the GEOV3 FVC for non-forest types (Figure 13(b)). Apart from the algorithmic differences between the NNs and the method of gap probability theory or the dimidiate pixel model, the differences caused by the scaling transformations cannot be ignored.

Figure 13. The scatter plots of FVC products over BELMANIP2 and ground measurement sites in China from 2018 to 2020. (a) Forest; (b) Non-forest.

5. Discussion

In comparison with two existing researches that estimated FVC based on GF-1 WFV images at regional scales (Jia et al. 2016; Li et al. 2016), one is the estimation of FVC based on the automated Monte Carlo unmixing analysis (AutoMCU) and SMA method from GF-1 WFV data (Li et al. 2016), and the other is the estimation of FVC based on a training backpropagation NNs using the simulations based on PROSAIL from GF-1 WFV (Jia et al. 2016). These two methods showed relatively reasonable results at regional scales, i.e. the accuracy of the AutoMCU method evaluated by ground measurements in Otindag Sandy Land was R²= 0.49 and RMSE = 0.17 and that of the SMA method was R²= 0.47 and RMSE = 0.27 (Li et al. 2016); and the accuracy of the NNs method was R²= 0.79 and RMSE = 0.07 using all ground measurements in Weichang County, Hebei Province, China (Jia et al. 2016). However, these two types of methods showed limitations when applied to large scales, i.e. the accuracy of the reflectance of the component pixels in both AutoMCU and SMA methods, the uncertainties arising from the uncontrolled interaction process of training NNs, and the possibilities of PROSAIL input parameter combinations for model training, and require the high performance of computational efficiency. The method proposed in this study, which can automatically and efficiently estimate the FVC with NDVI and LAI results on a large scale, has been successful in generating the FVC product with 16 m spatial resolution and 10-day temporal resolution over China. Moreover, the accuracy of GF-1 FVC, which has an RMSE of 0.15 and BIAS of 0.01, meets the requirements of vegetation monitoring according to direct evaluation with ground-based measurements.

According to the temporal consistency analysis results of Figure 11, the inversion values of GF-1 FVC are slightly higher than those of GEOV3 FVC in the early growth and receding phases, and this higher trend of FVC is also observed in the comparison results between GF-1 and GEOV3 FVC products in Figure 13(b). This is partly due to the sensitivity of the GF-1 FVC values upscaled from the 16 m spatial resolution sub-pixels during vegetation growth. In addition, the different estimation methods of the two products also led to inversion variance, the dimidiate pixel model method based on NDVI is sensitive to plant greening, while the NN trained from the PROSAIL model represents the comprehensive information of each vegetation type. For the direct evaluation between GF-1 and GEOV3 FVC products with 300 m spatial resolution in Figure 12, the GEOV3 FVC showed an obvious underestimation for grass type (most of FVC values less than 0.8) for Hulunbuir ground measurements, while GF-1 FVC showed a better agreement with FVC ground measurements. The large difference is caused by the mixed-pixels at 300 m, where the mean FVC value upscaled from 16 m to 300 m is closer to the ground measurements. Moreover, from a report of the scientific quality assessment of LAI, FPAR and FCOVER Collection 300 m Version 1 product, the values of GEOV3 FVC for grass type are less than 0.8 with the ground observations for the periods from 2014 to 2018 (Martínez-Sánchez, 2020). In this case, the GF-1 FVC estimation method proposed in this study is more suitable for monitoring the mid-temperate continental grassland types such as Hulunbuir, especially during the plant flowering period.

6. Conclusions

This study introduces an automated approach for producing a 16 m/10-day GF-1 FVC product that spans the years 2018–2020 in China. The method proposed in this research incorporates the gap probability theory method, which considers the clumping effect at both the canopy and landscape levels for forest types. Additionally, it combines the dimidiate pixel model method, which utilizes the greenness sensitivity of NDVI for non-forest types. The GF-1 FVC product demonstrated a level of accuracy (RMSE = 0.15) that was comparable to the GEOV3 FVC product (RMSE = 0.25) at a spatial resolution of 300 m when compared to FVC ground measurements at 7 sites across China. However, the GF-1 FVC product offers a higher spatial resolution of 16 m. The comparison results from 2019 revealed that there were stronger agreements between the GF-1 and GEOV FVC products for shrub and grass types in both January and July. In contrast, a significant difference (>0.2) was observed for broad-leaf forest and coniferous forest types in January. However, the temporal patterns exhibited by the aggregated GF-1 data show a lesser degree of smoothness in comparison to GEOV3. The presence of missing data continues to pose a substantial challenge, resulting in unreliable GF-1 FVC values, particularly for forest types. This issue arises due to adverse weather conditions and lower proportions of effective LAI estimation ratios derived from the gap probability theory approach. Consequently, enhancing the temporal consistency of GF-1 FVC is imperative for monitoring vegetation growth at a regional scale. Potential future endeavors may involve the utilization of GF-6, Sentinel-2, and other relevant resources to enhance the accessibility of input data or use spatio-temporal reconstruction algorithms to fill in the gaps caused by missing values.

Acknowledgments

The authors express their gratitude to all providers of ground measurements. The ground measurements in Beibei and Pu'er were obtained from satellite, aircraft and ground-based comprehensive experiments of authenticity testing for the large-scale special project of a high-resolution earth observation system in Beibei, Chongqing Municipality and Pu'er, Yunnan Province. The ground measurements of the Hulunbuir steppe in the Inner Mongolia Autonomous Region were taken by the Command Center of Natural Resource Comprehensive Survey. The ground measurements in the middle reaches of the Heihe River Basin, Gansu Province were obtained from the Cold and Arid Regions Environmental and Engineering Research Institute, Chinese Academy of Sciences (CAS). The ground measurements of the Saihanba plantation forestry farm in Hebei Province were provided by Chengde Remote Sensing Test Site, State Key Laboratory of Remote Sensing Sciences. The ground measurements of Huailai station in Hebei Province and Jingyuetan station in Jilin Province were provided by Professor Wen Jianguang and Bai Junhua of the Aerospace Information Research Institute, CAS, respectively.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The dataset of 16 m/10-day NDVI, LAI and FVC product of MuSyQ GF-series from 2018 to 2020 across China (version 01) can be freely downloaded from science data bank at https://www.scidb.cn/en (https://doi.org/10.11922/sciencedb.j00001.00252, https://doi.org/10.11922/sciencedb.j00001.00262 and https://doi.org/10.11922/science-db.j00001.00266). The world map of the Köppen-Geiger climate classification can be downloaded at http://koeppen-geiger.vu-wien.ac.at/present.htm. Both of the fine land-cover maps with 30 m spatial resolution in 2015 and the clumping index can be downloaded at http://www.geodata.cn (https://doi.org/10.12041/geodata.4200772.ver1.db and https://doi.org/10.12041/geodata.2127518316-57979.ver1.db). The GEOV3 FVC product can be freely downloaded from the CGLS website (https://land.copernicus.eu/global/products/ FCover). The BELMANIP-2.1 and Chinese CERN data that support the inter-comparisons of this study are available at https://calvalportal.ceos.org/web/olive/site-description and http://www.cnern.org.cn/. Please contact us if you need additional materials.

Additional information

Funding

This work was supported by National Key Research and Development Program under Grant 2019YFE0126700, and the National Natural Science Foundation of China under Grant 41871265.