Influence of Vegetation Phenology on the Temporal Effect of Crop Fractional Vegetation Cover Derived from Moderate-Resolution Imaging Spectroradiometer Nadir Bidirectional Reflectance Distribution Function–Adjusted Reflectance

Abstract

Moderate-Resolution Imaging Spectroradiometer (MODIS) Nadir Bidirectional Reflectance Distribution Function (BRDF)-Adjusted Reflectance (NBAR) products are being increasingly used for the quantitative remote sensing of vegetation. However, the assumption underlying the MODIS NBAR product’s inversion model—that surface anisotropy remains unchanged over the 16-day retrieval period—may be unreliable, especially since the canopy structure of vegetation undergoes stark changes at the start of season (SOS) and the end of season (EOS). Therefore, to investigate the MODIS NBAR product’s temporal effect on the quantitative remote sensing of crops at different stages of the growing seasons, this study selected typical phenological parameters, namely SOS, EOS, and the intervening stable growth of season (SGOS). The PROBA-V bioGEOphysical product Version 3 (GEOV3) Fractional Vegetation Cover (FVC) served as verification data, and the Pearson correlation coefficient (PCC) was used to compare and analyze the retrieval accuracy of FVC derived from the MODIS NBAR product and MODIS Surface Reflectance product. The Anisotropic Flat Index (AFX) was further employed to explore the influence of vegetation type and mixed pixel distribution characteristics on the BRDF shape under different stages of the growing seasons and different FVC; that was then combined with an NDVI spatial distribution map to assess the feasibility of using the reflectance of other characteristic directions besides NBAR for FVC correction. The results revealed the following: (1) Generally, at the SOSs and EOSs, the differences in PCCs before vs. after the NBAR correction mainly ranged from 0 to 0.1. This implies that the accuracy of FVC derived from MODIS NBAR is lower than that derived from MODIS Surface Reflectance. Conversely, during the SGOSs, the differences in PCCs before vs. after the NBAR correction ranged between –0.2 and 0, suggesting the accuracy of FVC derived from MODIS NBAR surpasses that derived from MODIS Surface Reflectance. (2) As vegetation phenology shifts, the ensuing differences in NDVI patterning and AFX can offer auxiliary information for enhanced vegetation classification and interpretation of mixed pixel distribution characteristics, which, when combined with NDVI at characteristic directional reflectance, could enable the accurate retrieval of FVC. Our results provide data support for the BRDF correction timescale effect of various stages of the growing seasons, highlighting the potential importance of considering how they differentially influence the temporal effect of NBAR corrections prior to monitoring vegetation when using the MODIS NBAR product.

1. Introduction

With rapid advances in multi-angle observation satellite-based sensors, such as Polarization and Directionality of the Earth’s Reflectances, Multi-angle Imaging SpectroRadiometer, and Moderate-Resolution Imaging Spectroradiometer (MODIS), whose detection bands extend beyond the visible and near-infrared (NIR) to the thermal infrared, researchers can now obtain richer spectral and structural information, which greatly improves the capability of multi-angle optical remote sensing in modeling and inversion [1]. The solar radiation reflected by most ground objects is anisotropic, with directionality varying according to solar incidence and the geometric conditions of satellite observations. The Bidirectional Reflectance Distribution Function (BRDF) is a key model for multi-angle optical remote sensing studies that is often used to characterize such changes. Studying the surface BRDF is significant for robust progress in the quantitative remote sensing of vegetation, which can describe the three-dimensional structural characteristics of ground objects and retrieve various surface parameters based on it, such as the Leaf Area Index and Fractional Vegetation Cover (FVC) [2]. Furthermore, bidirectional reflectance characteristics of the ground surface may be used to obtain the reflectance normalized to different angles, which can eliminate the reflectance discrepancies caused by distinct incidence or observation angles, paving the way for subsequent quantitative remote sensing applications [3]. Due to its advantages of a broad spectral range coupled with high temporal and spatial resolution, MODIS is widely used in surface cover detection, vegetation growth monitoring, and vegetation phenology change research [4,5]. The MODIS Nadir BRDF-Adjusted Reflectance (NBAR) product effectively adjusts the impact introduced by the BRDF effect and is commonly used for detecting shifts in vegetation phenology [6,7,8].

The MODIS NBAR product has a 16-day retrieval period, and its inversion model assumes that surface anisotropy remains unchanged during this period. Yet this assumption may be unreliable for areas that undergo rapid surface changes, such as floods or fires [9]. Moreover, various vegetation types such as crops, forests, shrubs, and grasses exhibit distinctive phenological characteristics, and the canopy structure of each vegetation type will change significantly at the start of season (SOS) and again at the end of season (EOS), which may also undermine that assumption’s reliability. Accordingly, Proud et al. proposed a modified version of the MODIS BRDF algorithm for the Spinning Enhanced Visible and InfraRed Imager data aboard the geostationary Meteosat Second Generation (MSG) satellites, which markedly shortened the retrieval period and enabled the usage of MSG BRDF data to examine short-term and rapidly evolving phenomena [9]. To investigate the appropriate retrieval period of the EVI2 time series after the BRDF correction of the Advanced Baseline Imager (ABI) data onboard the Geostationary Operational Environmental Satellite-R series satellites, Shen et al. used NBAR data of the Visible Infrared Imaging Radiometer Suite (VIIRS) for evaluation; they found that the ABI data retrieval period was shortened to 3 days, which is crucial for near-real-time monitoring of crop phenology vis-à-vis the 16-day retrieval period of the polar-orbiting satellite [10]. Hasegawa et al. constructed a BRDF observation system for forests to determine the relationship between vegetation phenology and BRDF, reporting a greater BRDF influence in summer [11]. However, that study did not explore whether the influence of the BRDF correction of quantitative remote sensing retrieval of vegetation at different stages of the growing seasons. Overall, the body of research addressing the relationship between vegetation phenology and BRDF has mostly focused on vegetation types such as forests, woodlands, and pastures [12,13,14], leaving crops vastly understudied.

As a pivotal indicator of surface vegetation conditions, FVC and its alteration—one of the main manifestations of land cover change—indirectly conveys the degree of human activities’ impact on natural ecology [15]. As the backbone of food production, cropland not only ensures the food security of a country but also is a critical ecosystem component, and shifts in FVC could substantially affect agricultural production and the stability of surface ecosystems. Common methods for retrieving fractional vegetation coverage from remote sensing images include the dimidiate pixel model and machine learning algorithms [16,17,18]. Due to the ±55° across-track scanning angle of MODIS, its images exhibit a more pronounced BRDF effect. Hence, to better apply MODIS NBAR product to dynamically monitor the growth status of crops, it is imperative to better understand in depth the temporal effect of the FVC derived from MODIS NBAR at different stages of the crops’ growing seasons.

The overarching goal of this study was to investigate the temporal effect of MODIS NBAR-derived FVC across distinct stages of the crops’ growing seasons. Specifically, the NDVI and EVI fitting curves of the MOD13Q1 product were first utilized to acquire the calendar date of crop phenological parameters, namely SOS and EOS as well as their intervening stable growth of season (SGOS). Next, we used the MODIS Surface Reflectance product (MOD09GA) and MODIS NBAR product (MCD43A4) to derive the FVC at the calendar date of those phenological parameters. Finally, using the PROBA-V bioGEOphysical product Version 3 (GEOV3) FVC as the verification data, the FVC’s retrieval accuracy was compared and analyzed to explore the temporal effect of the MODIS NBAR correction at each stage of the crops’ growing seasons.

2. Materials and Methods

2.1. Study Area

The study area is located in the Wancheng District, Nanyang City, Henan Province (China), encompassing 112°28′9″–112°49′52″ E and 32°37′51″–33°9′45″ N, in the middle of the Nanyang Basin. Lying in the transition between the subtropics and warm temperate zone, this region has a humid climate, with an average annual temperature of 14.9 °C and an average annual precipitation of 805.8 mm, which is conducive to the growth of various plants. The crop maturity system follows an annual double-cropping pattern, with wheat, corn, and peanuts predominately cultivated. The land cover types, sampling points distribution, and true-color images of the three mixed pixels in the study area are shown in Figure 1. The crop phenology extraction sampling points were used to subsequently acquire the NDVI/EVI average fitting curve to extract phenological parameters, with the BRDF sampling points (savanna pixel, grass pixel, crop pixel, and mixed pixels) then used to analyze the BRDF patterns of various land cover types.

Temperature and precipitation are the primary factors driving phenological shifts, but their influence on vegetation phenology varies across regions [19]. Here, the study area’s precipitation and temperature data were collected from the ERA5-Land monthly averaged product for 2017–2021 (Figure 2). Overall, the temperature from 2017 through 2021 showed clear seasonal fluctuations, whose pattern was generally similar during the period; precipitation also displayed seasonal fluctuations, with relatively low precipitation in the summer of 2019, which may have led to the advancement of stages of the crops’ growing seasons [20].

2.2. Data

The remote sensing datasets used in this study are listed in

Table 1. Remote sensing datasets used in this study.

Datasets	Temporal Resolution	Spatial Resolution	Purpose
MODIS vegetation indices product (MOD13Q1 v061)	16 days	250 m	To extract crop phenological parameters
MODIS Surface Reflectance product (MOD09GA v061)	daily	500 m	To provide surface reflectance and solar zenith angle
MODIS NBAR product (MCD43A4 v061)	daily	500 m	To provide NBAR reflectance
MODIS Land Cover product (MCD12Q1 v061)	yearly	500 m	To obtain cropland pixels
MODIS BRDF Model Parameters product (MCD43A1 v061)	daily	500 m	To provide RTLSR model kernel parameters
GEOV3 FVC product	10 days	300 m	To serve as validation data

. MOD13Q1 [21] is a MODIS vegetation indices product. Since that data have been processed already (with respect to water bodies, cloud cover, and heavy aerosols), their quality is guaranteed, spurring widespread use in studies of vegetation cover change [22,23]. We selected the NDVI and EVI time series from which to extract the crop phenological parameters of the study area. MOD09GA [24] is a MODIS Surface Reflectance product corrected for atmospheric conditions (such as gas, aerosol, and Rayleigh scattering), used here to retrieve the FVC without BRDF correction. MCD43A4 [25] is a MODIS NBAR product, whose view zenith angle is normalized to the nadir, while its solar zenith angle is adjusted to the local solar noon of the day to eliminate the impact of varying view zenith angles, it used in this study to retrieve the NBAR-corrected FVC.

The GEOV3 FVC product [26] is an FVC product generated by filtering, gap filling, and smoothing data based on a neural network model [27]. Compared with GEOV1 and GEOV2, the GEOV3 product has superior spatial resolution via algorithmic improvements. Several studies have evaluated its performance, showing that it performs well and has reliable accuracy [27,28,29]; hence, it was used in this study to evaluate MODIS-derived FVC. For comparative purposes, ArcMap was used to reproject and resample the GEOV3 FVC data to match the coordinate system and spatial resolution of MODIS data.

MCD12Q1 [30] is a MODIS land cover type product, which provides land cover classification information on a global scale. It is now one of the most widely used land cover products, containing a total of 13 scientific datasets, of which the IGBP classification scheme was used here. MCD43A1 [31] is a MODIS BRDF Model Parameters product, which supplies weighting parameters associated with the semi-empirical RTLSR (Ross Thick-LiSparse Reciprocal) BRDF model.

2.3. Methodology

This study sought to explore the temporal effect of MODIS NBAR correction at different stages of the growing seasons, which entailed three major processes: crop phenological parameters extraction, FVC retrieval, and accuracy evaluation. Their data processing flow is presented in Figure 3. Firstly, TIMESAT was used to fit the MOD13Q1-NDVI and MOD13Q1-EVI time series, resulting in the average fitting curve used to extract crop phenological parameters. Next, MOD09GA and MCD43A4 data collected under clear weather conditions were selected according to the previously acquired crop phenological parameters, then NDVI was calculated through band math, and the respective FVC estimated by using the dimidiate pixel model. Finally, the GEOV3 FVC product was used as verification data to compare and analyze the retrieval accuracy of the FVC derived from MOD09GA and MCD43A4.

2.3.1. Crop Phenological Parameters’ Extraction

TIMESAT (v3.3) is recognized for being a robust, efficient software tool for phenology studies. With its graphical user interface (GUI), TIMESAT is user-friendly in which key parameters can be easily set [32]. The SOS and EOS refer to the point at which the curve intersects a proportion of the seasonal amplitude as measured from the left minimum and right minimum, respectively. However, because the results of the first and last year’s phenological extraction were insufficiently accurate when using TIMESAT, the MOD13Q1 data from 2017 to 2022 were instead selected to extract crop phenological parameters from 2018 to 2021. MODIS-NDVI and MODIS-EVI are widely employed in regional or global phenology studies given their notable advantages, e.g., a high temporal and spatial resolution [12,33]. In the land cover type map, a total of 30 sampling points were selected across the study area’s cropland (Figure 1a) with reference to the Sentinel-2 true-color image, after which the average fitting curve of NDVI and EVI were obtained to extract the crop phenological parameters. The average value was taken as the final result to bolster the accuracy of phenological extraction. To reconstruct time series data, S-G (Savitzky–Golay) filtering is commonly used [34,35]. Following the suggestions of Jonsson et al. [36], who originally proposed the dynamic threshold, and by referring to other scholars’ experiences [37,38], we decided on setting the dynamic threshold to a value of 0.2, whose calculation formula is defined as follows:

$${\mathrm{VI}}_{\lim }=\left({\mathrm{VI}}_{\max }–{\mathrm{VI}}_{\min }\right)\mathrm{S}$$

(1)

where VI_lim is the extraction threshold of SOS (EOS) NDVI/EVI; VI_max is the maximum value of NDVI/EVI in the vegetation growth cycle; VI_min is the minimum value of NDVI/EVI in the rising or falling stage of that growth cycle; and S is the percentage coefficient, set to 20% here.

Lacking ground-based phenological observations, this study relied on the time series curve filtering method and dynamic threshold—given their good phenological extraction accuracy as demonstrated in previous studies [34,35,37,38]—and compared the phenological extraction results of two different vegetation indices: NDVI and EVI. Finally, their average value of crop phenological parameters was selected as the result. Based on the ground field survey of Ma et al. and known field details of the study area, the SOS of overwintering crops usually spans the end of October to the end of November [39]. Our other phenological results were consistent with those found relevant in the study area in previous years [37,38]. Further, our other phenological results for 2018 and 2019 showed high consistency with the phenological dataset from 2000 to 2019, released by Luo et al. [40].

2.3.2. Fractional Vegetation Cover (FVC) Retrieval

The phenological parameters derived from the two vegetation indices consisted of SOS, EOS, and SGOS. According to these, the MOD09GA and MCD43A4 data with clear weather near them were identified via the Google Earth Engine to retrieve the FVC. Next, MODIS Reprojection Tools and The Environment for Visualizing Images software (v5.3) were used to preprocess the NIR and RED bands data for the NDVI calculation. Then, based on this NDVI, the FVC was estimated by the dimidiate pixel model, which is the most widely used and simple linear mixture decomposition model. It assumes that each pixel’s information consists of vegetation as well as no vegetation, which greatly weakens the influence of soil background, atmosphere, and vegetation type, making it an effective method for retrieving FVC [17,41]. The FVC calculation formula is as follows:

$$\mathrm{FVC}=\left(\mathrm{NDVI}-{\mathrm{NDVI}}_{\mathrm{soil}}\right)/\left({\mathrm{NDVI}}_{\mathrm{veg}}-{\mathrm{NDVI}}_{\mathrm{soil}}\right)$$

(2)

where NDVI_soil denotes the NDVI value of bare soil or no vegetation coverage area, and NDVI_veg denotes the NDVI value of a pure vegetation coverage pixel. In this study, we counted the NDVI and used those NDVI values with cumulative probabilities of 5% and 95% to represent NDVI_soil and NDVI_veg, respectively.

2.3.3. Data Analysis

This section describes the methods employed to assess the FVC’s accuracy and the BRDF’s shape. The linear correlation between the MODIS-derived FVC and GEOV3 FVC was quantified using the Pearson correlation coefficient (PCC) to explore the temporal effect of the MODIS NBAR correction. The Anisotropic Flat Index (AFX) was used to assess the shape change in BRDF and thereby determine the relationship between geometric–optical scattering and volumetric scattering characteristics.

(1) FVC Accuracy Assessment: GEOV3 FVC was used to evaluate MODIS-derived FVC. Although the two datasets are not exactly the same, differing in terms of their (i) spatial resolution and spectral responses, as well as (ii) FVC retrieval algorithms, the GEOV3 FVC can still provide a sound reference for the MODIS-derived FVC. The all-cropland pixels in the study area were obtained using MCD12Q1, and the FVC derived from MOD09GA and MCD43A4 in Section 2.3.2 was statistically analyzed vis-à-vis GEOV3 FVC by the PCC, to explore the temporal effect of MODIS NBAR correction for crops at the different stages of the growing seasons. Specifically, the PCC was used to determine the magnitude of consistency between the MODIS-derived FVC and GEOV3 FVC. The PCC is calculated this way:

$$\mathrm{PCC}={\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{FVC}}_{\mathrm{MODIS}}^{\mathrm{i}}–\overline{{\mathrm{FVC}}_{\mathrm{MODIS}}}\right)\left({\mathrm{FVC}}_{\mathrm{GEOV}3}^{\mathrm{i}}–\overline{{\mathrm{FVC}}_{\mathrm{GEOV}3}}\right)}{\sqrt{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{FVC}}_{\mathrm{MODIS}}^{\mathrm{i}}–\overline{{\mathrm{FVC}}_{\mathrm{MODIS}}}\right)}^2}\sqrt{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{FVC}}_{\mathrm{GEOV}3}^{\mathrm{i}}–\overline{{\mathrm{FVC}}_{\mathrm{GEOV}3}}\right)}^2}}}$$

(3)

where ${\mathrm{FVC}}_{\mathrm{MODIS}}^{\mathrm{i}}$ is the FVC derived from MOD09GA or MCD43A4; the ${\mathrm{FVC}}_{\mathrm{GEOV}3}^{\mathrm{i}}$ is the FVC provided by GEOV3 FVC; and n is the number of all cropland pixels in the study area.

(2) BRDF shape assessment: Unlike a spectral vegetation index, such as NDVI, the MODIS BRDF shape indicator AFX contains pertinent information about vegetation structure. AFX conveys the BRDF shape based on its value around unity. An AFX < 1 indicates that geometric–optical scattering is dominant, implying a dome-shaped BRDF curve with a prominent reflectance peak (hotspot) in the retro-solar direction. An AFX > 1 indicates that volumetric scattering is dominant, implying a bowl-shaped BRDF curve where reflectance near the nadir is lower than for larger scattering angles, with the minimum usually displaced towards the forward scattering direction in the principal plane. An AFX ≈ 1 indicates a relatively flat BRDF curve. Importantly, AFX varies by band and phenological date, and is related to land cover type along with land surface structure. The general AFX formula is as follows [42]:

$$\mathrm{AFX}=1+{\displaystyle \frac{{\mathrm{f}}_{\mathrm{vol}}\left(\mathsf{\lambda }\right)}{{\mathrm{f}}_{\mathrm{iso}}\left(\mathsf{\lambda }\right)}}\times 0.189184–{\displaystyle \frac{{\mathrm{f}}_{\mathrm{geo}}\left(\mathsf{\lambda }\right)}{{\mathrm{f}}_{\mathrm{iso}}\left(\mathsf{\lambda }\right)}}\times 1.377622$$

(4)

where ${\mathrm{f}}_{\mathrm{iso}}\left(\mathsf{\lambda }\right)$, ${\mathrm{f}}_{\mathrm{geo}}\left(\mathsf{\lambda }\right)$, and ${\mathrm{f}}_{\mathrm{v}\mathrm{ol}}\left(\mathsf{\lambda }\right)$ are constant coefficients that, respectively, represent the weights of isotropic scattering, geometric-optical scattering, and volumetric scattering. The values 0.189184 and 1.377622 are the bi-hemispherical integral of the Ross Thick kernel and the reciprocal LiSparse kernel, respectively.

3. Results

3.1. Extraction of Phenological Parameters via NDVI and EVI

The fitted curves of the preprocessed NDVI and EVI time series generated by TIMESAT, along with phenological parameters, are shown in Figure 4. Evidently, the NDVI and EVI time series curves follow the same general fluctuation pattern, depicting similar seasonal variation in crop phenology; however, the EVI profile is always positioned below the NDVI profile. Meanwhile, as Figure 4 shows, their extracted SOS points lie close to each other, and likewise for EOS. After a comparative consideration, the midpoint of the NDVI and EVI phenological timing was taken and this designated as the crop phenological parameters results for 2018–2021. The annual phenological parameters of the NDVI time series curves in Figure 4 included two SOSs and EOSs, as well as four SGOSs. However, factors such as cloud cover prevented the identification of the phenological parameter SGOS during the rise in NDVI in the second growing season of 2020.

The extraction results of SOS and EOS in 2018–2021 are presented in

Table 2. The SOS (start of season) and EOS (end of season) of crop phenological parameters, from 2018 to 2021.

Year	SOS1 1	EOS1 1	SOS2 1	EOS2 1
2018	5 February 2018	14 May 2018	17 June 2018	22 September 2018
2019	18 November 2018	17 May 2019	20 June2019	15 September 2019
2020	25 October 2019	8 May 2020	19 June2020	24 September 2020
2021	18 November 2020	16 May 2021	18 June2021	30 September 2021

¹ SOS₁ and EOS₁ refer to crop phenological parameters in the first growing season; SOS₂ and EOS₂ refer to crop phenological parameters in the second growing season.

. The study area follows an annual double-cropping pattern, resulting in two growing seasons per year. Evidently, the first SOS in 2018 differed substantially from that in the other three years. Our investigation revealed that, from 28 August to 19 October 2017, Henan Province experienced its most serious sustained rainy weather in the last decade. This situation is evinced in Figure 2, where precipitation rose abnormally during this period, which severely impacted the harvesting of the autumn crop and strongly delayed the sowing of the overwintering crop. Hence, we speculated that could be why the first SOS in 2018 came so late.

Table 2 also indicates that the second EOS in 2019, along with both the SOS and EOS in the first growing season of 2020, happened earlier than in other years. Based on our investigation, this was mainly because the average temperature in the study area was above normal with precipitation less than normal since the sowing of the autumn crop in June 2019. This reasoning is also supported by the trends in Figure 2, which led to an earlier-than-normal harvest of the summer crop followed by the subsequent winter wheat being sown sooner than normal.

3.2. Overview of Fractional Vegetation Cover Estimation

After determining the calendar date of crop phenological parameters from the time series fitted curves shown in Figure 2, and given that cloud cover will reduce the quality and availability of remote sensing images, we selected MODIS data with clear-sky weather near the calendar date of phenological parameters. This calendar date information on typical phenological points in 2021 is provided, as an example, in

Table 3. Typical crop phenological parameters’ temporal information (2021, example).

MODIS Data: Calendar Date of Phenology	GEOV3 Data: Calendar Date of FVC
15 November 2020 (SOS)	20 November 2020
10 February 2021 (SGOS)	10 February 2021
9 May 2021 (SGOS)	10 May 2021
22 May 2021 (EOS)	20 May 2021
22 June 2021 (SOS)	20 June 2021
30 July 2021 (SGOS)	31 July 2021
8 September 2021 (SGOS)	10 September 2021
30 September 2021 (EOS)	30 September 2021

(these corresponding to those points in Figure 4).

From Table 3, three distinct phenological parameters—15 November 2020 (SOS), 10 February 2021 (SGOS), and 30 September 2021 (EOS)—were selected as examples. The FVC was derived from MOD09GA and MCD43A4, and the visualized differences in FVC values generated accordingly (Figure 5). These results revealed that the spatial distribution of FVC derived from MOD09GA and MCD43A4 is generally similar (Figure 5a,b,d,e,g,h, respectively). Some blank (white) data did appear in the middle of the left side in Figure 5b,e,h, and this is due to a river in the area, which led to missing data in the MCD43A4 product. According to Figure 5c,f,i, in general, the difference between MOD09GA and MCD43A4 is greater than zero. The difference in FVC between Figure 5c,i (SOS and EOS, respectively) is notably higher than that in Figure 5f (SGOS).

3.3. Effects of the NBAR Correction at Different Stages of the Growing Seasons on Fractional Vegetation Cover Retrieval Results

The Pearson correlation coefficients of MOD09GA and MCD43A4 with the GEOV3 FVC, respectively, and their differences are shown in Figure 6. The Pearson correlation coefficients, PCC_MOD09GA and PCC_MCD43A4, between FVC_MOD09GA, FVC_MCD43A4, and GEOV3 FVC at selected phenological parameters from 2018 through 2021 are plotted in Figure 6a. These phenological parameters correspond to those identified by the NDVI average fitting curve shown in Figure 4. From Figure 6a, it can be seen how the temporal effect of the retrieved FVC varies before and after NBAR correction at different stages of the growing seasons. The reason for the breakpoint in 2020 was mentioned in Section 3.1.

The relationship between the differences in PCC_MOD09GA and PCC_MCD43A4 is depicted in Figure 6b. When the difference is >0, this would indicate that the retrieval accuracy of the FVC derived before the NBAR correction exceeds that after the NBAR correction; and vice versa when the difference is <0. Overall, Figure 6b shows that, during SGOS of the crop, the differences between PCC_MOD09GA and PCC_MCD43A4 ranged between –0.2 and 0, indicating a generally stronger correlation between FVC_MCD43A4 and GEOV3 FVC. This result suggested the FVC derived from MCD43A4 (NBAR) was more consistent with GEOV3 FVC. However, for most SOSs and EOSs, the differences between PCC_MOD09GA and PCC_MCD43A4 mainly ranged between 0 and 0.1. This implied a weaker correlation between FVC_MCD43A4 and GEOV3 FVC, demonstrating that despite NBAR correction, the MCD43A4 product was less consistent with GEOV3 FVC than FVC_MOD09GA was with GEOV3 FVC.

4. Discussion

4.1. Effects of Different Stages of the Growing Seasons on the Temporal Effect of Fractional Vegetation Cover Retrieved from MODIS NBAR

Normalizing the multi-angular reflectances or spectral indices to a common solar and view angle using BRDF models has been demonstrated to effectively mediate the BRDF impacts, thus enabling a more accurate intercomparison of satellite observations, both spatially and temporally [9,43]. Therefore, the MCD43A4 product is now widely used to study vegetation phenology. Common BRDF models, such as the Roujean model, the Walthall model, and the RTLSR model, are designed particularly for polar-orbiting satellite observations conducted under a wide range of angular conditions (MODIS, VIIRS, etc.) [44]. But high-quality BRDF modeling requires a sufficient number of clear-sky observations within a short time span, such that no significant changes in vegetation properties have occurred. The polar-orbiting satellite retrieval time period usually spans multiple days, typical 8–16 days, but it could be unreliable for regions known to undergo rapid land surface variations, such as fires and floods [9]. For some of the rapid developmental stages in vegetation phenology, vegetation characteristics will change, albeit not as drastically as when fires or floods occur, yet the model assumption of surface anisotropy remains unchanged over the retrieval period may still be unreliable.

As Figure 6b shows, most of the correlations between the FVC derived from the MCD43A4 product and the GEOV3 FVC are weaker when the crop is at SOS or EOS, which will impair the monitoring of crop phenology by MODIS NBAR. This negative impact is primarily due to the fact that MODIS is a polar-orbiting satellite sensor with a 16-day retrieval period for the NBAR product, for which the model assumption that the surface anisotropy remains unchanged during the period imposes an intrinsic uncertainty, especially during key stages of the growing seasons such as SOS and EOS when vegetation characteristics undergo rapid alterations.

4.2. Analysis of the Influence of Vegetation Types on the BRDF for Different Stages of the Growing Seasons

To explore the influence of different vegetation types on BRDF patterns in the different stages of the growing seasons, this study obtained the MCD43A1 product of the three vegetation type pixels with different stages of the growing seasons shown in Figure 1a to create NDVI polarization maps. These new results are shown in Figure 7. The data are formatted in a polar coordinate system, with the polar angle representing the relative azimuth angle (ranging from 0° to 360° with an interval of 1°) and the distance from the center of the circle representing the viewing zenith angle (ranging from 0° to 70° with an interval of 0.5°). The backscattering direction corresponds to an azimuth angle of 0°, while the forward scattering direction corresponds to an azimuth angle of 180°. Satellite observations with a high sun/view zenith angle (>70°) will reduce accuracy in the atmospheric correction process and further negatively influence BRDF modeling [45], so the view zenith angle is limited here to between 0° and 70°. To comprehensively consider the different BRDF patterns of the NIR and RED bands, we used the RTLSR model to simulate both bands and thereby generate an NDVI spatial distribution map. The SOS and EOS of crop phenological parameters in 2021 are employed as examples for analysis.

From the perspective of NDVI patterns, we find that various vegetation types exhibit distinct patterns in their NDVI at different stages of the growing seasons. Figure 7a–d show that winter and summer crops feature differential NDVI patterns at SOS and EOS. The shape of the high-value area in the NDVI pattern of winter crop undergoes a more pronounced change, whereas for summer crop, it changes less markedly. In Figure 7e–h, with shifts in vegetation phenology, the shape of the high-value area in the NDVI pattern of savanna pixel changes less, indicating it is less affected by seasonal changes. In contrast Figure 7i–l reveal that for a grassland pixel, the shape of the high-value area in the NDVI pattern changes more noticeably with shifts in vegetation phenology. Figure 7i,l also show that NDVI patterns differ starkly from those found for crop pixel and savanna pixel.

From the perspective of AFX, as vegetation phenology shifts, there is a growth and decline in vegetation, which would affect the relative size relationship between volumetric and geometric–optical scattering. Meanwhile, the NIR and RED bands have distinct characteristics for each vegetation type in the different stages of the growing seasons. For example, on 15 November 2020, for crop pixel the AFX_RED = 0.734 and AFX_NIR = 1.226 (Figure 7a), indicating a dome-shaped BRDF for the RED band, but a bowl-shaped for the NIR band; for savanna pixel the AFX_RED = 0.737 and AFX_NIR = 0.888 (Figure 7e), indicating a dome-shaped BRDF for both bands; for grassland pixel the AFX_RED = 1.023 and AFX_NIR = 1.292, likewise indicating a bowl-shaped BRDF for both bands.

NDVI is arguably the most commonly used remote sensing index to quantify FVC [46]. Here, the FVC values in the figures come from the GEOV3 FVC product. Figure 7 shows that the values of NDVI vary with characteristic directional reflectance (i.e., reflectance under the corresponding characteristic direction), while the values of FVC are not the same in different stages of the growing seasons. The differences between the NDVI patterns and AFX reflect the complex influence of different vegetation types and seasonal changes on the BRDF patterns, which can provide valuable auxiliary information for the accurate identification and classification of vegetation types. By combining the NDVI under the characteristic directional reflectance, it becomes possible to accurately estimate the FVC of different vegetation types.

4.3. Analysis of the Influence of Mixed Pixel Distribution Characteristics on the BRDF for Different Stages of the Growing Seasons

To explore the influence of the distribution characteristics of the mixed pixels on the BRDF patterns in the different stages of the growing seasons, this study obtained the MCD43A1 product of the three mixed pixels with different stages of the growing seasons in Figure 1a, to create NDVI polarization maps. The distribution characteristics of the three mixed pixels are shown in Figure 1b–d. The results are shown in Figure 8. The format for drawing the NDVI polarization map is the same as that in Section 4.2.

From the perspective of NDVI patterns, the distribution characteristics of different mixed pixels can evidently affect the patterns of NDVI, with the mixing of buildings and crops leading to greater pattern complexity. Figure 8a–d show that the direction of the high-value area in these NDVI patterns undergoes a reversal, from the backscattering direction to the forward scattering direction as the vegetation phenology shifts. Figure 8e–h show that the shape of the high-value area in these NDVI patterns is altered more markedly by the change shifts in vegetation phenology. Figure 8i–l show that with shifts in vegetation phenology, there is a more pronounced modification of the shape and direction of the high-value area in these NDVI patterns. Also, as seen in Figure 8a,b,e,f under the same stage of the growing seasons, the high-value areas of these NDVI patterns for the two sets of mixed pixels are distributed in opposite directions. This outcome may be caused by the opposing distribution of characteristics in mixed pixels.

From the perspective of AFX, as vegetation phenology shifts, crops in mixed pixels undergo growth and decline, altering the relative size relationship between the volumetric and geometric–optical scattering. Meanwhile, we find that NIR and RED bands display different characteristics for the mixed pixels in different stages of the growing seasons. For example, on 15 November 2020, Figure 8a shows an AFX_RED = 0.911 and AFX_NIR = 0.874, indicating a dome-shaped BRDF for either band; Figure 8e shows an AFX_RED = 0.934 and AFX_NIR = 1.067, indicating a dome-shaped BRDF for the RED band, yet a bowl-shaped for NIR band; Figure 8i shows an AFX_RED = 0.960 and AFX_NIR = 0.902, indicating a dome-shaped BRDF for both bands.

Figure 8 also depicts how the NDVI values vary with characteristic directional reflectance, with the FVC values obviously changing in different stages of the growing seasons. The differences between the NDVI patterns and AFX reflect the complex influence of mixed pixels’ distribution characteristics and seasonal changes on the BRDF patterns, which can provide valuable auxiliary information for the accurate interpretation of not only the distribution characteristics but also coverage ratio of the mixed pixels. So, by combining the NDVI under the characteristic directional reflectance could provide an opportunity for the accurate estimation of the FVC for mixed pixels.

4.4. Limitations and Future Prospects

By studying the temporal effect of the MODIS NBAR product when crops are in different stages of the growing seasons, we have gained a better understanding of the relationship between the MODIS NBAR product and crop phenology, which should enable a more accurate monitoring of vegetation. Nonetheless, it is worthwhile to consider this study’s limitations. (1) Different crop types exhibit varying spectral characteristics, growth cycles, and structural characteristics, which can lead to discrepancies in vegetation index time series curves. As the study area expands, the diversity of crop types in the same period increases; this could reduce accuracy of phenological identification, thereby enlarging the uncertainty of this study’s results. Additionally, significantly expanding the study area makes it harder to acquire MODIS data that meets observation conditions because of factors such as cloud cover. For those reasons, we restricted the study area to the Wancheng District.

(2) The dimidiate pixel model has the advantages of a simple principle and strong operability, but it also harbors certain disadvantages. In particular, the selection of NDVI values for bare soil and pure vegetation pixels is affected by subjective factors. At the same time, it is inappropriate to divide mixed pixels into bare soil and photosynthetic vegetation in areas covered by extensive non-photosynthetic vegetation. However, these shortcomings do not outweigh the model’s benefits, which also include straightforward, reliable calculations; universal and easy-to-obtain data parameters; and high retrieval accuracy. Therefore, it is still widely used in FVC retrieval [17,41]. (3) Discrepancies between the phenological periods acquired at the pixel scale from remote sensing data and those recorded at the species level from ground observations, can lead to significant uncertainties in the validation of phenological extraction results based on remote sensing data [47]. Lacking ground-based phenological observations, this study employed robust extraction methods with high phenological periods’ accuracy in previous studies. Indirect methods were then applied to validate the phenology extraction results, demonstrating a good correlation [37,38,40]. Consequently, the vegetation phenology derived from the MODIS vegetation indices product can be considered to have a certain degree of reliability.

High-quality BRDF modeling requires not only obtaining enough clear-sky observations within a short time span, but also considering the representativeness of angular sampling [48,49]. Polar-orbiting satellites have a longer retrieval period, but this allows for more representative angular sampling. Geostationary satellites can obtain more clear-sky observations in a short-term period with a large spread of illumination conditions. Still, the inherently invariant viewing geometry of the geostationary platform reduces the anisotropic variations within an observed reflectance time series. Considering the high complementarities of geostationary and polar-orbiting satellite observations, their robust integration will offer new opportunities for vegetation monitoring. In tandem, their integration creates opportunities to enhance the accuracy of remote sensing extraction of vegetation phenology.

5. Conclusions

This study examined the Wancheng District from 2018 to 2021, comparing the correlation between the crop FVC (Fractional Vegetation Cover) derived from MOD09GA and MCD43A4 with the GEOV3 FVC to quantitatively assess the temporal effect of MODIS NBAR across different stages of the growing seasons. The results demonstrate that using MODIS NBAR to retrieve crop FVC is less effective at SOS or EOS, which poses a challenge for accurate vegetation dynamics monitoring. This study improves our understanding of the relationship between MODIS NBAR and vegetation phenology and provides a basis for more accurate monitoring of crop phenology using MODIS data in the future. Meanwhile, under different stages of the growing seasons, the differences in NDVI patterns and AFX can provide auxiliary information useful for vegetation classification and interpretation of mixed pixel distribution characteristics. Combining this with the NDVI at characteristic directional reflectance should bolster the accurate retrieval of FVC. Further study is warranted regarding the accuracy and applicability of MODIS NBAR in monitoring other vegetation types in different regions and various phenological conditions, as well as investigating the influence of characteristic directional reflectance on the temporal effect of the MODIS FVC correction. Pursuing such work will augment the scope of MODIS data’s application in quantitative remote sensing products for vegetation and will be the focus of our future research efforts.