GF-1 WFV Surface Reflectance Quality Evaluation in Countries along “the Belt and Road”

Abstract

The GaoFen-1 wide field of view (GF-1 WFV) has produced level 1 digital number data globally; however, most applications have focused on China, and data quality outside China has not been validated. This study presents a preliminary assessment of the 2020 GF-1 WFV surface reflectance data for Nepal, Azerbaijan, Kenya, and Sri Lanka along “the Belt and Road” route using Sentinel-2 Multi-Spectral Instrument (MSI), Landsat-8 Operational Land Image (OLI), and Moderate Resolution Imaging Spectroradiometer (MODIS) data. A method for obtaining the GF-1 WFV surface reflectance data was also proposed, with steps including atmospheric correction, cross-radiation calibration, and bidirectional reflectance distribution function correction. The results showed that WFV surface reflectance data was not significantly different from MSI, OLI, and MODIS surface reflectance data. In the visible and near-infrared bands, for most landcover types, the bias was less than 0.02, and the precision and root mean square error were less than 0.04. When the landcover types were forest and water, the MSI, OLI, and MODIS surface reflectance data were higher than that of WFV in the near-infrared band. The results of this study provide a basis for assessing the global application potential of GF-1 WFV.

1. Introduction

The GaoFen-1 (GF-1) satellite is an important component of China’s high-definition Earth observation satellite series. GF-1 is equipped with four wide field of view (WFV) sensors, each with a swath of up to 200 km, and a combined four camera swath of nearly 800 km, covering four wavelengths from visible to near-infrared (NIR) with a spatial resolution of 16 m and a temporal resolution of only 4 days [1]. The combination of high spatiotemporal resolution and wide coverage enables GF-1 images to be used in a variety of applications such as seasonal crop classification mapping [2,3], water resource environmental monitoring [4,5,6], leaf area index assessment [7,8], vegetation cover estimation [9,10,11], and high-resolution aerosol optical depth retrieval [12,13,14,15]. The GF series satellites already play a major role in China’s high-precision earth observation, and there is already usable analysis-ready data (ARD) for the GF-1 WFV; however, only the processing data within China has been completed, while fewer studies have been conducted on the processing of data outside China with the quality evaluation of surface reflectance [16]. As GF data become more sophisticated, we need to pay more attention to the potential of the GF series of satellites to provide high-quality services on a global scale.

The GF-1 is equipped with four sensors, WFV1–4. WFV2 and WFV3 observe zenith angles between 0° and 24°, whereas WFV1 and WFV4 observe zenith angles between 24° and 40°. The larger observed zenith angle and increase in atmospheric paths can cause a dramatic difference in reflectance between the zenith and off-zenith views [17,18]. In addition to the normal geometric and atmospheric corrections, the surface reflectance inversion of the GF-1 WFV also needs to take into consideration the bidirectional reflectance distribution function correction (BRDF) effect over a large field of view. Fortunately, the Moderate Resolution Imaging Spectroradiometer (MODIS) has developed validated global BRDF products that can be applied to simulate the BRDF effects produced by non-Lambertian surface targets in the direction of the observation object of interest [19,20,21]. No application examples of BRDF corrections for GF-1 WFV using MODIS products are available at present.

Surface reflectance is the ratio of the radiation reflected from a surface to the incident radiation, characterizing the ability of a surface to absorb and reflect solar radiation [22]. It is a fundamental remote sensing parameter for quantitative remote sensing, and its precision determines the quality and accuracy of applications such as surface parameter inversion and feature classification [23,24]. A variety of sensors similar to the GF-1 WFV are available, and these sensors already have surface reflectance data that is directly available to researchers, such as Landsat and Sentinel, and the quality of the data is widely recognized worldwide. For example, the United States Geological Survey developed an ARD for the Landsat series of satellites. The ARD is provided as a tiled, georegistered, top of atmosphere (TOA) and surface reflectance, defined in a common equal-area projection, with quality assessment. The processed ARD is currently available for Landsat-5 Thematic Mapper, Landsat-7 Enhanced Thematic Mapper Plus, Landsat-8 Operational Land Imager (OLI), and Thermal Infrared Sensor data [25]. Studies have been conducted to evaluate the quality of the Landsat-5, 7, and 8 series globally, such as cross-validation using MODIS data or validation using surface reflectance inverted from Aerosol Robotic Network (AERONET) data. The validation results show that, in more than 80% of the scenarios, the surface reflectance uncertainty of Landsat-5, 7 remains within specifications, defined as 0.071 × surface reflectance + 0.0071. The validation results for Landsat-8 are superior to those for Landsat-5 and 7 [26,27,28,29]. For Sentinel-2 multi-spectral instrument (MSI) data, users only need to perform simple processing to obtain surface reflectance data, which is facilitated by the atmospheric correction plug-in Sen2Cor developed by the European Space Agency [30,31]. Sentinel-2 surface reflectance data processed with Sen2Cor was compared with atmospherically corrected Landsat-8 data, and the overall RMSE was 0.031 for all six bands used for comparison [32]. The surface reflectance data of Gaofen-1 has not been validated outside China, compared with the quality of the same type of satellite products that have been validated worldwide.

To address the abovementioned problems, we compared the surface reflectance data of GF-1 WFV with that of Sentinel-2 MSI, Landsat-8 OLI and MODIS for quality assessment and proposed a BRDF correction method for GF-1 WFV. The main contributions of this article can be summarized follows:

• This study validates the quality of GF-1 WFV surface reflectance data from Nepal, Azerbaijan, Kenya, and Sri Lanka. It provides quality information for the global application of GF-1 WFV and quality references for harmonization with Sentinel-2 MSI and Landsat-8 OLI.
• A nadir-view surface reflectance processing flow for WFV is proposed, including geometric corrections, atmospheric corrections, and BRDF corrections to reduce the wide field of view effect of the sensor.

2. Study Area

Nepal, Azerbaijan, Kenya, and Sri Lanka were selected for the GF-1 WFV data quality evaluation, and two experimental areas were selected for each country, as shown in Figure 1. A zoomed-in view of the land cover types in each experimental area is shown in Figure 2. The details of each area are listed in

Table 1. Surface cover types and corresponding GF-1 WFV imaging dates for each study area.

Area ID	Country	Landcover Type	GF-1 WFV Imaging Date
1	Nepal	grassland with some mixed evergreen forest	27 October 2020
2	Nepal	impermeable surface in the central area and forested grassland at the edges	30 October 2020
3	Azerbaijan	grassland	19 July 2020
4	Azerbaijan	agricultural land	19 July 2020
5	Kenya	dominated by sparse shrubs with some grassland	17 January 2020
6	Kenya	bare ground with essentially no vegetation cover	17 January 2020
7	Sri Lanka	forested grassland	11 December 2020
8	Sri Lanka	mixed savanna and agricultural land	11 December 2020

, each covering approximately 2500 km². These four countries have a wide range of landcover types and climate categories and provide sufficient experimental samples to verify the surface reflectance accuracy of the GF-1 WFV [33,34,35,36].

3. Data and Methodology

3.1. GF-1 WFV Satellite Data and Auxiliary Data

The GF-1 WFV data were provided by the China Center for Resources Satellite Data and Application (CRESDA), which currently releases GF-1 data at level 1 for download at https://data.cresda.cn/#/home (accessed on 26 September 2023). The downloaded GF-1 WFV images contained blue, green, and red visible bands and an NIR band, at a spatial resolution of 16 m. This was accompanied by solar and satellite angular information at the time of imaging and the rational polynomial coefficients (RPCs) used for orthographic correction. The thumbnails of the GF-1 WFV images for the eight experimental areas are shown in Figure 3; all images were composited in true color. The selected experimental areas have excellent image quality, rich landcover types, and no obvious cloud cover.

Auxiliary data included elevation, aerosol optical depth (AOD), and BRDF data. Shuttle Radar Topography Mission-SRTM 1 Arc-Second Global was selected as the digital elevation model (DEM) for orthorectification with a spatial resolution of 30 m [37]. MCD19A2 provides AOD data, where MODIS has a transit time similar to that of GF-1 and negligible aerosol variability in the absence of clouds and can therefore be used as an AOD parameter for GF-1 atmospheric correction with a spatial resolution of 1 km [38]. The MCD43A1 data provided the weighting parameters associated with the BRDF model for correction of the GF-1 WFV data [39].

The products used for quality evaluation were Sentinel-2 MSI, Landsat-8 OLI, and MODIS surface reflectance data. The MSI and OLI surface reflectance were both atmospherically corrected for high accuracy and had spatial resolutions of 10 m and 30 m, respectively. The MODIS surface reflectance employs the MCD43A product, which is nadir BRDF-adjusted reflectance data with a spatial resolution of 500 m. All three datasets contained the blue, green, red, and NIR bands. MSI, OLI, and MODIS data were downloaded from the Google Earth Engine (GEE). A summary of the remote sensing data information used in this article is shown in

Table 2. Remote sensing data information.

Data	Purpose of the Data	Spatial Resolution (m)	Source of Data
GF-1 WFV	surface reflectance inversion	16	CRESDA
MCD19A2	Provide atmospheric calibration parameter AOD	500	GEE
MCD43A1	Provide BRDF calibration parameters	500	GEE
SRTM 1	DEM	30	USGS
Sentinel-2 MSI	Surface reflectance quality evaluation reference data	10	GEE
Landsat-8 OLI	Surface reflectance quality evaluation reference data	30	GEE
MCD43A	Surface reflectance quality evaluation reference data	500	GEE

.

3.2. Inversion Method for GF-1 WFV Surface Reflectance

This study presents a GF-1 WFV surface reflectance inversion process that solves the geometric, radiometric, and spectral aspects of the problem, including orthorectification, radiometric calibration, geometric correction, atmospheric correction, cross-radiation calibration, and BRDF correction. The overall data processing flow is shown in Figure 4 and described in detail in each section.

3.2.1. TOA Reflectance Calculations and Orthographic Correction

The process of generating surface reflectance requires the completion of the radiometric calibration of the GF-1WFV image, which is the calculation of the image from digital number values to radiance [40,41]. Once the radiance is obtained, the TOA can be calculated according to radiative transfer theory. The calibration factors and solar radiance were obtained from the GF-1 satellite parameters published by CRESDA. The GF-1 WFV level 1 data require orthographic correction so that the pixels can accurately represent the actual ground position. CRESDA distributes level 1 data with RPCs to facilitate complete user orthorectification according to the RPC model [42,43]. To improve the orthographic correction accuracy, this study adopted the Landsat-8 OLI B8 panchromatic band as the digital orthophoto reference, provided the required ground control points, and adopted a DEM with a 30 m spatial resolution as the elevation reference.

3.2.2. Cross-Radiation Calibrations and Atmospheric Correction

GF-1 has four sensors: WFV1, WFV2, WFV3, and WFV4. Despite the similar external conditions at the time of imaging, there are some differences in the sensor configurations that require cross-radiation calibrations normalized to the same standard. When all four sensors are imaged simultaneously, there is an overlapping area characterized by a minor variation in the geometric angle information and an almost uniform external environment [44,45]. In this study, we utilized this feature to determine the cross-radiation calibration coefficients by a linear regression of the image values in the overlapping areas of different sensors [46,47]. The cross-radiation calibration of the GF-1 data was based on the WFV2 sensor and corrected for the other sensors.

The second simulation of the satellite signal in the solar spectrum (6S) radiative transfer model was used to complete the atmospheric correction of the GF-1 WFV. The model is based on mathematical and physical theories and presents the solar radiation transfer process in the solar–earth–satellite system in a scientific and reasonable manner, with a clear physical meaning [22]. The necessary inputs for the atmospheric correction of the 6S model, such as aerosol data and satellite geometry, are available.

Aerosol data used MODIS aerosol types and AOD data (MCD19A2), and MODIS has transit times similar to GF-1; solar angle and observation angle data are incidental to GF-1 WFV. According to the radiative transfer equation of the 6S model, the TOA reflectance can be generally expressed as shown in Equation (1), under the usual assumption of a homogeneous Lambertian surface for the simulation of the feature:

$${\rho }_{TOA}\left({\theta }_s,{\theta }_i,\varphi \right)={\rho }_a\left({\theta }_s,{\theta }_i,\varphi \right)+\frac{{\rho }_{BOA}}{1-{\rho }_{BOA}\cdot S}T\left({\theta }_s\right)T\left({\theta }_i\right)$$

(1)

where, ${\theta }_s$, ${\theta }_i$, $\varphi$ are the solar zenith angle, the observation zenith angle, and the solar azimuth relative to the sensor; ${\rho }_a$ is the course radiation due to Rayleigh scattering and aerosol scattering; ${\rho }_{BOA}$ is the surface reflectance; $S$ is the atmospheric spherical reflectance; $T({\theta }_s)$ is the total atmospheric transmittance from the sun to the feature; and $T({\theta }_i)$ is the total atmospheric transmittance from the feature to the satellite.

3.2.3. BRDF Calibration

The reflection of incident solar radiation from the Earth’s surface is isotropic, while the GF-1 WFV single-scene image has a width of up to 200 km and a satellite zenith angle of up to approximately 40°, with a difference of up to 20° in the satellite zenith angle [1]. Together with variations in the solar zenith angle and relative azimuth, this can lead to large differences in the reflectance of the same feature in multi-scene images acquired at adjacent times.

The MCD43A1 product provides the weighting parameters ($f_{iso}$, $f_{vol}$, $f_{geo}$) associated with the BRDF model for each band. The bi-directional reflectivity of any illumination angle of incidence and observation angle can be derived from the extrapolation or interpolation of the kernel-driven model, as shown in Equation (2). The volume scattering kernel chosen for this study is the Ross thick kernel, and the geometrical optical scattering kernel is the Li sparse reciprocal kernel [48,49,50].

$$R({\theta }_s,{\theta }_i,\varphi ;\gamma )=f_{\begin{array}{l}iso\\ \end{array}}\left(\gamma \right)+f_{vol}\left(\gamma \right)\times k_{vol}\left({\theta }_s,{\theta }_i,\varphi \right)+f_{geo}\left(\gamma \right)\times k_{geo}\left({\theta }_s,{\theta }_i,\varphi \right)$$

(2)

$R$ is the reflectance at solar zenith angle ${\theta }_s$, observed zenith angle ${\theta }_i$, and relative azimuth angle $\varphi$; $k_{vol}$ is the volume scattering kernel, and $k_{geo}$ is the geometrical optics kernel, which are functions of incidence and observation angles; $f_{iso}$, $f_{vol}$, and $f_{geo}$ are constants representing the ratio of the isotropic, volume scattering, and geometrical optics components of reflectance, respectively.

The processing method is to first generate an a priori BRDF model library for the study area for one year based on the MCD43A1 data. Based on the kernel-driven model, the simulated reflectance $R^{\prime }$ with the same observation geometry as the actual reflectance $R$, and the zenith-direction simulated reflectance $R_n$ are calculated using forward simulation with a priori BRDF knowledge. Based on the relationship between $R$, $R^{\prime }$ and $R_n^{\prime }$, the zenith-direction reflectance $R_n$ is calculated.

$$R_n=R/R^{\prime }\times R_n^{\prime }$$

(3)

This completes the preprocessing of the GF-1 WFV data to obtain the WFV surface reflectance in the near-zenith direction.

3.3. Evaluation of GF-1 WFV Surface Reflectance

3.3.1. Selection Criteria for Sentinel-2 MSI, Landsat-8 OLI and MODIS Data

In this study, MSI, OLI, and MODIS surface reflectances were selected to evaluate the quality of the WFV surface reflectance obtained from the processing described above. The criterion for selecting MSI and OLI surface reflectance is that the image with the lowest amount of clouds is preferred with imaging dates no more than seven days before or after WFV surface reflectance; if no suitable image was available within seven days before or after, the search was extended to find an image with the most similar imaging time. The MODIS surface reflectance product is an average product generated every 16 d; therefore, it was selected for the same day as the WFV surface reflectance. For the processing of different spatial resolutions, the upsampling method was selected, that is, in the WFV comparison with MSI, the MSI was aggregated to 16 m to be consistent with WFV; in the WFV comparison with OLI, the WFV was aggregated to 30 m for consistency with OLI; in the WFV comparison with MODIS, the WFV was aggregated to 500 m to be consistent with MODIS, and the resampling methods were all selected for bilinear interpolation. The effect of clouds and cloud shadows in the images was removed, and reflectance values less than 0 and greater than 0.95 were considered outliers.

3.3.2. Spectral Adjustment

The waveband settings for each sensor are shown in Figure 5, which shows that there is still a difference in the visible-NIR band width between the WFV and the MSI, OLI, and MODIS sensors, which is the major reason for the difference in the surface reflectance. Spectral adjustment of MSI, OLI, and MODIS surface reflectance is required to reduce surface reflectance differences due to changes in sensors and ensure that the accuracy check reflects the true variance between sensors [51,52].

The following steps were adopted for specific spectral adjustment: first, the simulated reflectance of the features in the spectral library was calculated based on the spectral response functions of different sensors; second, the simulated reflectance of all features under different relative spectral responses was linearly regressed to establish the band conversion factor. The spectral adjustment factors are shown in Figure 6. The hyperspectral feature library used in this study contained surface reflectance spectral samples with 5637 features [53]. As mentioned earlier, the four WFV sensors were cross-calibrated with the WFV2 standard. Therefore, band conversion was performed using the coefficients corresponding to the relative spectral response of WFV2. In addition, the difference between the Terra and Aqua MODIS sensors and the difference between sensors A and B of Sentinel-2 was evaluated.

3.3.3. Comparison Metrics

To assess the quality of the WFV surface reflectance, three statistical indicators–bias, precision, and root mean square error (RMSE)–were applied to quantify the deviation of the WFV surface reflectance from the MSI, OLI, and MODIS reference data. The bias, precision, and RMSE values represent the variance between the comparison datasets, with lower values representing better matches, where bias corresponds to the average deviation between the two datasets, and precision corresponds to the change in the average deviation [27]. Bias, precision, and RMSE were calculated using the specific formulae shown in Equations (4)–(6):

$$bias=\frac{1}{n}\times {\displaystyle \sum _{i=1}^n{\epsilon }_i}$$

(4)

$$precision=\sqrt{\frac{1}{n-1}\times {{\displaystyle \sum _{i=1}^n\left({\epsilon }_i-bias\right)}}^2}$$

(5)

$$RMSE=\sqrt{\frac{1}{n}\times {{\displaystyle \sum _{i=1}^n{\epsilon }_i^2}}^{}}$$

(6)

where $n$ is the number of valid samples used for the comparison and $\epsilon$ is the surface reflectance of the WFV minus the surface reflectance of the reference sensors.

4. Results

4.1. Satellite Image Selection Results

The WFV, MSI, OLI, and MODIS surface reflectance products used in this study are listed in

Table 3. Acquisition times for WFV, MSI, OLI, and MODIS images.

	Sensors	WFV	MSI	OLI	MODIS
Area ID
1		27 October 2020	25 October 2020	25 October 2020	27 October 2020
2		30 October 2020	1 November 2020	20 October 2020	30 October 2020
3		19 July 2020	20 July 2020	20 July 2020	19 July 2020
4		19 July 2020	17 July 2020	13 July 2020	19 July 2020
5		17 January 2020	17 January 2020	17 January 2020	17 January 2020
6		17 January 2020	22 January 2020	17 January 2020	17 January 2020
7		11 December 2020	11 December 2020	7 December 2020	11 December 2020
8		11 December 2020	13 December 2020	7 December 2020	11 December 2020

. The image thumbnails are shown in Figure 7, all images are true color composite images. The absence of MSI and OLI images is due to the fact that the images are at the edges of the imaging trajectory or due to a cloud block effect.

4.2. Evaluation Results of WFV and MSI Surface Reflectance

In this section, we calculated the statistical indicators for the MSI and WFV surface reflectance and analyzed the differences between the two images. Figure 8 shows the scatter plots of the MSI and WFV surface reflectance in each area, as well as the statistical indicator bias, precision, and RMSE. WFV and MSI comparisons were performed at a spatial resolution of 16 m, with spectral conversion, geometric matching, and the removal of cloud-affected and heavily missing Areas 1, 4, and 8 prior to comparison.

Among the various landcover types, MSI and WFV surface reflectance in the visible-NIR band bias was below 0.02, precision and RMSE were below 0.04, and the difference between the two images was not significant. In the NIR band, when the landcover type was forest, MSI surface reflectance was higher than WFV surface reflectance, with significant differences, and the statistical indicators were sharply higher.

The Area 3 landcover type was mainly grassland, Area 5 was dominated by sparse shrubs with some grassland, and Area 6 was bare ground with essentially no vegetation cover. In these three areas, the bias in the visible-NIR band was basically below 0.02, and precision and RMSE were below 0.03. The lower bias and RMSE values indicated that the absolute difference between the MSI and WFV surface reflectance in these areas was not significant. Lower precision indicates a more stable range in variation between the two datasets. A special case occurred in the red bands of Areas 3 and 6, where both precision and RMSE increased slightly to 0.03. This may have been due to the presence of distinctly demarcated landcover types in the two areas, which require more geometrical alignment and result in an increase in precision and RMSE, which describe the range of variation in pixel value deviation.

Area 2 is an urban built-up area with mostly impervious surfaces in the center and forests at the edges. Area 7 has a landcover type of forest grassland. The scatter plots of the two areas showed that the bias, precision, and RMSE of the MSI and WFV surface reflectance were significantly lower in the blue, green, and red bands than in the NIR band, whereas the statistical indicators in the NIR band appeared to increase sharply. The NIR scatter plot showed that a portion of the MSI surface reflectance was significantly higher than the WFV surface reflectance. Viewing the image pixels corresponding to the landcover type, the image pixels with deviations in Area 2 had a mixed forest landcover type, and the image pixels with deviations in Area 7 had a broad-leaf evergreen forest landcover type.

4.3. Evaluation Results of WFV and OLI Surface Reflectance

In this section, we calculated the statistical indicators for the OLI and WFV surface reflectance and analyze the differences between the two images. Figure 9 shows the scatter plots of the OLI and WFV surface reflectance in each area, as well as the statistical indicator bias, precision, and RMSE. WFV and OLI comparisons were performed at a spatial resolution of 30 m, with spectral conversion, geometric matching, and removal of cloud-affected and heavily missing Areas 2, 4, and 7, prior to the comparison.

The results showed that in most landcover types, the bias in the visible-NIR band of OLI and WFV surface reflectance was below 0.02, with precision and RMSE below 0.03. When the landcover type was forest and water, the NIR band image pixel values of OLI surface reflectance were higher than those of WFV surface reflectance, and the statistical indicators were significantly higher.

The Area 3, Area 5, and Area 6 surface cover types were mainly grassland, sparse shrubs with some grassland, and bare ground with essentially no vegetation cover, respectively. In these three areas, the bias in the visible-NIR band was below 0.02, and the precision and RMSE were below 0.025. The lower bias and RMSE values indicate that the absolute difference between the OLI and WFV surface reflectance in these areas was not significant. A lower precision indicates a more stable range of variation between the two.

The landcover types in Area 1 were grassland, mixed evergreen forest, and bare ground while Area 8 was a mixed area of savanna and agricultural land. Scatter plots in the visible band showed low statistical indicators in both areas, with bias below 0.02, precision and RMSE around 0.03, and no significant difference between the OLI and WFV surface reflectance. However, the scatter plots in the NIR bands of Areas 1 and 8 showed large differences in the OLI and WFV surface reflectance. The precision and RMSE were high for both areas, indicating a wide range of image values. In the Area 1 image, the pixels with large differences in NIR bands represented a landcover type of evergreen forest while in the Area 8 image, it indicated a landcover type of water and evergreen broad-leaf forest. For these landcover types, the WFV surface reflectance was significantly lower than the OLI surface reflectance. Anomalies with large deviations in the OLI surface reflectance could also be observed in the visible band scatter plot for Area 8. This region had cloud fringes in the image leading to elevated statistical indicators.

4.4. Evaluation Results of WFV and MODIS Surface Reflectance

In this section, we calculated the statistical indicators for the MODIS and WFV surface reflectance and analyze the differences between the two images. Figure 10 shows the scatter plots of the MODIS and WFV surface reflectance in each area, as well as the statistical indicator bias, precision, and RMSE. The WFV and MODIS comparisons were performed at a spatial resolution of 500 m with spectral conversion and geometric matching prior to comparison.

As a result of our experiments, we concluded that for most landcover types, MODIS surface reflectance in the visible and NIR bands was not significantly different from that of the WFV surface reflectance, with deviations below 0.02, and precision and RMSE below 0.04. In the NIR band, when the landcover types were forest and water, the MODIS surface reflectance was significantly higher than the WFV surface reflectance, and there was a large difference between the two datasets. In the red band, the MODIS surface reflectance was slightly higher than the WFV surface reflectance when the landcover type was bare.

In Areas 3–6, for visible-NIR bands, the bias was less than 0.02, and precision and RMSE were approximately 0.03. The lower bias and RMSE values indicate that the absolute difference between MODIS and WFV surface reflectance in these areas was not significant. Lower precision indicates a more stable range in variation between the two. A special case occurred in the red band of Area 6, where the statistical indicators increased with a precision and RMSE of approximately 0.037. The MODIS surface reflectance was higher than the WFV surface reflectance, and an examination of the image revealed that the landcover type for these pixels was bare ground.

Scatter plots for Areas 1–2 and 7–8 show that the statistical indicator bias, precision, and RMSE of MODIS and WFV surface reflectance were significantly lower in the visible bands than in the NIR band, while the statistical indicators in the NIR band appeared to increase sharply. The scatter plot of the NIR band showed the points where the MODSI surface reflectance was significantly higher than the WFV surface reflectance. Examination of the imagery revealed that the landcover type for this portion of the image elements in Areas 1–2 was a mixed evergreen forest while in Areas 7–8 it was an evergreen broad-leaf forest. In the scatter plot of the Area 8 NIR band in Figure 10, there are some points where the WFV surface reflectance (<0.05) is much lower than that of the MODIS surface reflectance, and inspection reveals that the landcover type of these pixels is water, which is similar to the previous comparison with OLI.

4.5. Comparison of Evaluation Results

The bias, precision, and RMSE statistics for the WFV surface reflectance when compared with the results from the MSI, OLI, and MODIS, for each of the eight experimental areas, are shown in

Table 4. Summary of statistical indicators. Bolding indicates that the statistical indicator exceeds the standard.

	Indicators			Bias			Precision			RMSE
Area ID
			Sensors	MSI	OLI	MODIS	MSI	OLI	MODIS	MSI	OLI	MODIS
		Band
1		Blue		/	0.0	−0.008	/	0.017	0.019	/	0.017	0.021
		Green		/	−0.011	−0.008		0.024	0.024	/	0.026	0.026
		Red		/	−0.01	0.002		0.031	0.03	/	0.032	0.03
		NIR		/	−0.003	−0.014		0.046	0.049	/	0.046	0.051
2		Blue		0.011	/	−0.009	0.022	/	0.019	0.025	/	0.021
		Green		−0.003	/	−0.004	0.024	/	0.024	0.024	/	0.024
		Red		0.008	/	0.006	0.026	/	0.026	0.027	/	0.026
		NIR		0.007	/	−0.014	0.067	/	0.049	0.067	/	0.051
3		Blue		−0.004	0.005	−0.007	0.02	0.014	0.019	0.021	0.015	0.02
		Green		0.013	−0.005	0.003	0.021	0.017	0.022	0.024	0.018	0.022
		Red		0.021	−0.001	−0.002	0.027	0.022	0.029	0.034	0.022	0.029
		NIR		0.008	−0.01	0.006	0.025	0.02	0.026	0.026	0.022	0.026
4		Blue		/	/	0.019	/	/	0.018	/	/	0.026
		Green		/	/	0.015	/	/	0.021	/	/	0.026
		Red		/	/	0.003	/	/	0.031	/	/	0.031
		NIR		/	/	−0.004	/	/	0.038	/	/	0.038
5		Blue		0.002	0.002	−0.007	0.01	0.007	0.011	0.01	0.007	0.014
		Green		0.007	−0.005	0.003	0.011	0.008	0.015	0.013	0.009	0.016
		Red		0.009	0.004	−0.01	0.013	0.01	0.019	0.016	0.01	0.022
		NIR		−0.012	−0.016	−0.003	0.015	0.01	0.02	0.019	0.019	0.02
6		Blue		0.009	0.005	0.004	0.018	0.013	0.022	0.02	0.014	0.022
		Green		0	0.007	0.017	0.024	0.016	0.029	0.024	0.018	0.033
		Red		−0.001	0.002	0.002	0.03	0.02	0.037	0.03	0.02	0.037
		NIR		−0.001	−0.013	−0.001	0.018	0.013	0.025	0.018	0.018	0.025
7		Blue		0.026	/	0.007	0.01	/	0.012	0.028	/	0.014
		Green		0.015	/	0.008	0.01	/	0.013	0.018	/	0.015
		Red		0.006	/	−0.004	0.01	/	0.013	0.012	/	0.014
		NIR		−0.043	/	−0.055	0.033	/	0.031	0.054	/	0.063
8		Blue		/	0.026	−0.004	/	0.019	0.019	/	0.032	0.019
		Green		/	0.012	0.01	/	0.021	0.02	/	0.024	0.022
		Red		/	−0.002	−0.004	/	0.021	0.021	/	0.021	0.021
		NIR		/	−0.04	−0.055	/	0.034	0.037	/	0.053	0.067

and Figure 11. The results showed that WFV, compared to the three datasets, had a bias below 0.02, and precision and RMSE below 0.04 for most landcover types. There were no significant differences between WFV and the three datasets, except for the forest and water landcover types, where WFV had the lowest statistical indicators and the strongest correlation with OLI.

As shown in Figure 11, the bias of the blue band in MSI Area 7 and OLI Area 8 was significantly higher than 0.02. This is because both areas are in Sri Lanka, which is cloudy throughout the year. The imaging conditions were harsh, and there were still clouds or thicker aerosols, even in the case of careful selection. The effect of clouds or aerosols decreased as the wavelength increased, and the bias in Areas 7–8 shows a decreasing trend from blue to green to red. In addition, although the RMSE of MSI and OLI in the visible band of Areas 7–8 was below 0.04, which is not outstanding, this pattern was also observed. The bias and RMSE accurately described the difference between the two datasets.

From the results of the three comparisons, the NIR band bias and RMSE of Areas 7–8 were significantly higher than those of Areas 1–2, which may be related to the forest area. Examination of the images of Areas 1–2 and 7–8 showed that all four areas had different degrees of forest cover, in which Areas 1–2 had a lower degree of forest cover than Areas 7–8. The bias and RMSE were also related to the degree of forest cover, while Area 8 had additional influence from bodies of water. The anomaly in the comparison was the NIR band of MSI Area 2, which showed anomalous precision and RMSE. It is speculated that this is because the middle of Area 2 is the Kathmandu area and has more surface buildings; resampling the MSI to 16 m added more mixed surface types to the image pixels, thus introducing more variance. This can also be seen in the Area 2 scatter plot in Section 4.2, where the correlation between the two images decreased significantly compared to the other areas of MSI.

5. Discussion

The WFV surface reflectance was compared with the MSI, OLI, and MODIS surface reflectances, and the statistical indicator bias, precision, and RMSE were calculated (Figure 8, Figure 9 and Figure 10). The results showed that WFV surface reflectance was not significantly different from MSI, OLI, and MODIS surface reflectance. In the visible and NIR bands, for most landcover types, the bias was less than 0.02, and the precision and RMSE were less than 0.04. When the landcover types were forest and water, the MSI, OLI, and MODIS surface reflectances were higher than the WFV surface reflectance in the NIR band. In this section, we analyzed the reasons for the differences between WFV and the other three sensors.

5.1. Effects of Landcover Type

For most landcover types, the consistency of WFV with MSI, OLI, and MODIS is high, with significant differences only in the NIR bands when the surface is forested and contains bodies of water. Combined with the feature spectral curves and the relative spectral response of the WFV, we hold the following more detailed discussion. From the spectral data collected by the airborne visible infrared imaging spectrometer in the United States Geological Survey spectral library, seven plant categories were identified: willow, white bark, sagebrush, lodgepole-pine, grass-fescue-wheat, aspen, and Douglas-fir, as shown in Figure 12. Viewing the spectral curves of each type of plant, we found that between 675 nm and 725 nm (the part in the red box), there is a significant difference in the reflectance of grass, shrubs, wheat and various types of trees, with the reflectance of grass, shrubs and wheat in the range of 0.1–0.15, while various types of trees show a clear reflectance trough below 0.05. The effective range of the NIR band spectral response function of the GF-1 WFV starts at 700 nm, and a wide NIR band may cause the WFV sensor to be affected by the red-edge effect of plants. This may explain the large differences in reflectance between the WFV surface reflectance and the MSI, OLI, and MODIS surface reflectance for forest landcover types, whereas no significant differences were observed in other vegetation areas. It has also been mentioned by some researchers that the normalized difference vegetation index and fractional vegetation cover of the WFV inversion do not perform as well as other vegetation types in forest areas [54,55,56]. We validated GF-1 WFV with Sentinel-2 MSI, Landsat-8 OLI based on the IGBP feature classes in MCD12Q1. We only counted feature classes with pixel counts greater than 10,000 and added the results as an Appendix A. Because the spatial resolution of MCD12Q1 is 500 m, the results of the feature class validation are affected by this. We will discuss the specifics in future work.

The NIR bands of the MSI, OLI, and MODIS surface reflectances were significantly higher than the WFV surface reflectance when the landcover type was water. In the NIR band, the ability of water to absorb light gradually increases, leading to a relatively low reflectance of water in the NIR band and making the reflectance inversion of water inherently difficult [57]. Meanwhile, the OLI and MODIS are designed to narrow the spectral range in the NIR band by removing the effect of water vapor absorption at 0.825 μm, while the WFV still maintains a wide spectral range in the NIR band. Therefore, when the landcover type was water, the differences in surface reflectance between the WFV and MSI, OLI, and MODIS in the NIR band were large.

5.2. Effects of Residual Clouds, Cloud Shadows, and Inconsistent Sensor Imaging Times

Clouds have significantly higher pixel values than the features, and residual clouds can seriously affect the evaluation quality. In Figure 9 (WFV compared to OLI), some points where OLI is significantly higher than WFV appear in the visible band scatter plot of Area 8. A review of the corresponding images reveals that both OLI surface reflectances in Area 8 have significant cloud cover. This is because OLI surface reflectance and MSI surface reflectance are downloaded from GEE. Whereas de-clouding in GEE uses a relatively simple cloud masking method based only on the quality assessment bands, although the main extensive cloud masses are removed, significant blurred patches remain near the cloud masses that can affect the quality assessment [58,59]. In this study, to validate the WFV data, the WFV images were selected to be cloud-free. To ensure that MSI and OLI are close in time to the WFV and that the features do not change significantly, some clouds are present in the MSI and OLI images for the guaranteed time range.

Changes in features over time can cause errors in quality assessment, such as large changes in plant spectra during the plant growing season with a difference in imaging time of 1–2 weeks [60,61] or ephemeral changes in the ground that cause damage to features [62]. As the transit times of the three satellites were inconsistent, it was difficult to collect images with imaging times on the same day. Even if images at similar times are retrieved, they may be affected by clouds or thick aerosols and cannot be used. In this study, we tried to ensure that the imaging time was within one week, and only the OLI surface reflectance data from Area 2 was 10 days away from the WFV, but it was not used because of excessive clouds. Nevertheless, the differences between WFV and MSI, OLI, and MODIS remained insignificant in most cases, with a high degree of consistency, suggesting that similar transit times did not have a major impact on quality evaluation.

5.3. Effects of Geographic Alignment

The effect of the geographic alignment on the validation results was significant [63,64]. There is a deviation in the geo-location of the GF-1 WFV in mountainous areas, and this part of the deviation will have a large impact on the quality evaluation, which will affect the joint application of the GF-1 WFV with other similar sensors on a global scale. For example, in Area 2 of the WFV and MSI, the positional deviation of the two datasets before calibration was extremely large, as shown in Figure 13, selecting part of the images in Area 2. The validation results before and after alignment are shown in

Table 5. Statistical indicators before and after geographical alignment. Bolding indicates a higher statistical indicator.

	Indicators	Band	Bias	Precision	RMSE
Area ID
Area 2.a		Blue	−0.01	0.03	0.032
		Green	−0.003	0.033	0.034
		Red	0.008	0.038	0.039
		NIR	0.007	0.074	0.074
Area 2.b		Blue	−0.011	0.022	0.025
		Green	−0.003	0.024	0.024
		Red	0.008	0.026	0.027
		NIR	0.007	0.067	0.067

and Figure 14. Before alignment, the bias between the two images was below 0.01; however, the precision and RMSE were above 0.03, and the linear relationship between the two datasets was not obvious. After alignment, the bias remained basically unchanged, the precision decreased by 9% to 32%, the RMSE decreased by 9% to 31%, and the linear relationship between the two datasets improved significantly. Geographic alignment significantly affects the results of the quality assessment, particularly in areas with large differences in landcover types and distinct feature demarcation lines, resulting in an increase in precision and RMSE. This is because the bias describes the overall deviation of the two-scene image; therefore, it is largely unaffected by geographic alignment, whereas precision and RMSE describe the range of data variation.

5.4. Limitations and Future Prospects

In this study, WFV data were compared with MSI, OLI, and MODIS surface reflectance data to investigate the quality of the data outside China. Although it was demonstrated that there was no significant difference between the WFV and the three datasets for most landcover types, we still need to continuously improve the inversion accuracy of the WFV surface reflectance in our future studies. To validate the surface reflectance, it is best to collect spectral data in the field for experiments based on the WFV transit time. In addition, the accuracy of the geometric alignment of WFV outside China remains a problem to be solved. The NIR bands of WFV and other sensors differ greatly, and a simple linear spectral adjustment is not sufficient to eliminate the differences between WFV, OLI, and MODIS.

6. Conclusions

In this study, eight areas with typical landcover types, including Nepal, Azerbaijan, Kenya, and Sri Lanka, were selected to evaluate the surface reflectance data of GF-1 WFV outside China using Sentinel-2 MSI, Landsat-8 OLI, and MODIS surface reflectance data as references. We proposed a method for obtaining high-accuracy GF-1 WFV surface reflectance data, including radiometric calibration, geometric correction, atmospheric correction, and cross-radiation calibration. Considering the large observation angle range of the WFV, the BRDF correction of the WFV surface reflectance was completed with MCD43A1. Quality assessment considers the differences in sensor performance and begins with the spectral transformation of the MSI, OLI, and MOIDS surface reflectance. For differences in spatial resolution, we chose to compare WFV with MSI, OLI, and MODIS at 16 m, 30 m, and 500 m. To quantify the variation between the WFV surface reflectance and other data, the bias, precision, and RMSE were selected as the evaluation indicators in this study. The results of the study showed that there was a high level of agreement between the WFV and MSI, OLI, and MODIS surface reflectance data, with differences only when the landcover types were forest and water. The quality of reference images can have a large impact on quality evaluation. Furthermore, the GF-1 WFV data from mountainous areas outside of China were found to have large geographic alignment errors when compared with the MSI and OLI, which could affect the global application of the GF-1. The GF-1 WFV preprocessing method proposed in this study was able to obtain accurate and reliable surface reflectance data, and the quality of the GF-1 WFV data was verified outside China. The quality evaluation of the dataset improves the credibility of the WFV surface reflectance data, which can meet the needs of practical applications. At present, this provides a reliable basis for the global use of GF-1 WFV, which will make global users aware of the usefulness of Chinese satellite data.