1. Introduction
Accurate estimation of crop biophysical and biochemical parameters during crop growing season is important for improving crop field management [
1]. Two important indicators of these parameters-leaf area index (LAI) and above ground biomass (AGB)-were used to monitor crop canopy structural development and growth changes and to estimate yield. The reasonable and reliable estimation of LAI and biomass can improve crop fertilizer applications [
2], water irrigation [
3,
4], disease and weed control [
5,
6], and grain production marketing [
7,
8,
9]. LAI and biomass change seasonally under different environmental conditions, and therefore, it is important to timely estimate their values. These parameters are traditionally estimated through destructive, time-consuming
in situ methods, which are difficult to conduct when crops cover large regions.
Owing to its capacity to obtain information on global and regional scales, remote sensing has become an effective tool for estimating LAI and biomass over large areas. Crop canopy structure mainly affects the spectral reflectance of crop canopy in the near-infrared (NIR) and visible spectrums. Numerous studies have shown a strong correlation between vegetation indices (VIs) and LAI and biomass using different integrations of visible and NIR reflectance [
10,
11,
12,
13,
14,
15,
16,
17]. A previous study has shown that normalized difference vegetation index (NDVI) was very sensitive to low LAI values (
i.e., LAI < 3) and saturation exists at medium to high LAI values (
i.e., LAI > 3) [
16]. Similarly, the saturation of NDVI values was shown at medium to high values of fresh biomass (around 2000 g/m
2) [
13]. The simple ratio [
16], the modified triangular vegetation index 2 (MTVI2) [
18], and the cumulative MTVI2 [
19] have shown better sensitivity at medium to high LAI and biomass. Previous results have shown that VIs based on the reflectance of red-edge bands (e.g., the red-edge triangular vegetation index (RTVI) and the modified chlorophyll absorption ratio index (MCARI2)) have great potential for improving estimations of LAI and biomass [
13,
18]. Most VIs have mainly been derived from field spectral radiometers [
16,
17,
20], airborne spectrographic imagers [
18], medium resolution spectrometers [
20], and high-resolution spectrometers [
21]. However, optical satellite images often have some limitations with respect to VIs because of the saturation problem and the subsequent reduction in estimation accuracy at medium to high LAI and biomass [
13,
14,
15,
16,
17,
22].
Compared with optical satellite images, synthetic aperture radars (SARs) have some advantages for monitoring crop growth status at medium to high LAI and biomass owing to the fact that microwave sensors have longer wavelengths, can penetrate crop canopies, and are not influenced by the presence of clouds or haze [
23]. However, SAR images are limited by the technique’s imaging geometry and radiation mechanism [
24]. Several SARs have been launched, such as ALOS-PALSAR (Japan), TerraSAR-X (Germany), Sentinel 1 (European Space Agency), and Radarsat1 and 2 (Canada). Some SARs have a short revisit time and high spatial resolution, which could be beneficial for monitoring crop development and health status [
25,
26]. Many studies have estimated LAI and biomass based on SAR images data acquired from either airborne or space-borne platforms [
27,
28,
29,
30,
31]. Some studies have shown that SAR backscattering was well correlated with biomass, especially that characterized by medium fractional cover [
32,
33,
34]. Since optical and SAR image data respond to crop characteristics differently, their complementary information content can support the estimation of crop conditions [
12]. The combination of optical and SAR image data has been used for the estimation of the LAI and biomass of crops and forests, and the results have shown that the estimated values agree well with the actual values [
27,
35,
36,
37]. Gao
et al. estimated the LAI, height, and biomass of maize using single-temporal environment and disaster monitoring satellite constellation (Huanjing (HJ)-1A/B) and RADARSAT-2; the results showed that this integrated method of determining VIs were well correlated with the LAI, height, and biomass near the maize heading stage [
24]. However, few studies have combined the optical and SAR data based on multi-temporal images for estimating the LAI and biomass of winter wheat.
Winter wheat is a main crop in Shaanxi Province. The accurate estimation of LAI and biomass for this crop is important for agricultural management and production in this region. HJ-1A/B data provides ground surface spectral information at a 30-m spatial resolution with a two-day revisit frequency (see
Section 2.3.1). Compared with other satellite data, HJ-1A/B data is a very good solution to balance the problems of spatial and temporal resolution. Thus, the HJ-A/B data with high spatial and temporal resolutions can offer an opportunity to monitor winter wheat growth status efficiently and objectively over large areas. In this study, the integration of high resolution SAR (RADARSAT-2) and optical image data (HJ-A/B) based on multi-temporal images data was further used to boost the estimation power of the LAI and biomass of winter wheat without adding to the concept of optical-SAR fusion. The major objectives of this study were the following: (i) to investigate the relationships of LAI and biomass with several optical spectral vegetation indices (OSVIs) and radar polarimetric parameters (RPPs), (ii) to estimate LAI and biomass with combined OSVIs and RPPs (the product of OSVIs and RPPs (COSVI-RPPs)), and (iii) to test and compare multiple stepwise regression (MSR) and partial least squares regression (PLSR) methods for estimating and improving the estimation accuracy of LAI and biomass of winter wheat based on the OSVIs, RPPs, and COSVI-RPPs. This study provides a good guideline for winter wheat field management.
4. Discussion
In this study, six optical spectral vegetation indices (OSVIs) were used to analyze the relationships of OSVIs with LAI and biomass for estimating LAI and biomass in winter wheat. The results showed that OSVIs were correlated with LAI and biomass (
Table 6). The effects of LAI and biomass on crop canopy spectral reflectance in the NIR and visible spectrum are known [
10,
11,
12,
13,
14,
15,
16,
17]. Therefore, these OSVIs based on different combinations of visible and NIR reflectance was significantly related with LAI and biomass. The results of our study confirmed previous results [
13,
14,
15,
16,
17]. The results showed that MTVI2 and EVI more accurately estimate LAI and biomass than other OSVIs. In addition to the red and NIR bands, EVI includes the blue band, which was used to correct for aerosol influences in the red band. Furthermore, the EVI is an optimized index designed to enhance the vegetation signal with improved sensitivity in high biomass regions through decoupling of the canopy background signal and a reduction in atmospheric and soil background noise influences [
53]. Therefore, EVI improves the estimation accuracy of LAI and biomass. The MTVI2 includes the green, red, and NIR bands. The decrease or increase in these bands reflectance influences the total area of the triangle, which was highly related with LAI [
18]. In order to reduce soil contamination effects, a soil adjustment factor is incorporated into MTVI2. The results of Haboudane
et al. [
18] indicated that MTVI2 was more sensitive to medium–high LAI. Therefore, the MTVI2 was used to boost the estimation accuracy of LAI and biomass. The results indicated that the EVI and MTVI2 could be used to estimate LAI and biomass in winter wheat. The results of OSVIs showed that the least-correlated with LAI and biomass was RVI1. Our results were consistent with the study of Gao
et al. [
24]. In contrast, OSAVI, SAVI, and NDVI were very sensitive to low LAI (LAI < 3) and were saturated at medium to high LAI values (LAI > 3) [
16,
50,
51,
52]. In this study, most of the LAI values were higher than 3. Therefore, MTVI2 and EVI were better than OSAVI, SAVI, and NDVI for estimating LAI and biomass. These results suggested that the OSVIs could be used to estimate LAI and biomass in winter wheat.
Fifteen radar polarimetric parameters (RPPs) were used to analyze the relationships between LAI, biomass and RPPs. The results showed that good correlations existed with the exception of the Dbl/Span and Odd/Span. The RVI and DERD indices exhibited the strongest correlations with LAI and biomass (
Table 7). The results of Koay
et al. [
57] suggested that the increase in HH during the tillering to filling stages was the main reason for the increase in single-volume backscattering as rice canopy became much denser. However, the denser paddy plants canopy showed more vertically oriented scatter, which led to a gradual reduction in the VV from the tillering to filling stages. As for HV, the double-volume scattering is the dominant scattering source at four winter wheat growth stages. The RVI not only included HH, HV, VV, backscattering difference information and then was sensitive to crops structure, but also reduced the environmental and incidence angle effects [
48,
58]. Therefore, the RVI showed higher correlations with LAI and biomass. The DERD are derived from the eigen-decomposition of the coherency matrix considering the reflection symmetry hypothesis. The results of Allain
et al. [
47] indicated that DERD provides a better inversion of crop parameters in their natural environment because it is easier to discriminate the different scattering mechanisms and eliminate the additive noise term for reducing the biases over the sample eigenvalues. Hence DERD was highly correlated with LAI and biomass. The HH, HV, VV, HH/VV, HH/HV, and VV/HV indices were strongly correlated with LAI and biomass. Previous studies have found that polarization ratios (HH/VV, HH/HV, and VV/HV) and backscattering coefficients (HH, HV, and VV) are suitable for LAI and biomass estimations in some crops and forests [
9,
24,
36]. Our results were in agreement with these studies. As alpha, anisotropy, and entropy were used to identify the scattering type and its relevance [
46], these indices were well correlated with LAI and biomass. The reason Vol/Span was correlated with LAI and biomass was because LAI and biomass largely influenced the range of the Vol change. Gao
et al. [
24] also indicated that Vol had a strong relationship with LAI and biomass. Our results further confirmed theirs results. The values of Odd and Dbl changed little and irregularly in the study of Gao
et al. [
24] and in our results. Therefore, Odd/Span and Dbl/Span were not correlated with LAI or biomass. The results indicated that the most of the RPPs were suitable for estimating LAI and biomass in winter wheat.
Because the NIR reflectance was not sensitive to the LAI or biomass of winter wheat at medium to high LAI, most of the OSVIs demonstrated the saturation phenomenon. However, SAR has some advantages for estimating LAI and biomass at medium to high LAI and biomass [
58], and therefore, RPPs were introduced in our study in combination with multispectral data. Previous studies have combined OSVIs and RPPs to estimate biomass and LAI in crops or forests by simply multiplying them [
12,
24,
35,
36]. Their results indicated that this method can be used to improve the estimation accuracy of biomass and LAI. Therefore, we combined the optical spectral vegetation indices and radar polarimetric parameters to estimate biomass and LAI in winter wheat by simply multiplying them. The combined indices RVI × OSVIs and DERD × OSVIs were created based on the good relationships between LAI, biomass and OSVIs, RPPs. Compared with the OSVIs and RPPs alone, the results showed that the COSVI-RPPs were more suitable to estimate biomass and LAI at medium to high vegetation coverage. The values of R
2 were 0.68 for LAI and 0.80 for biomass, respectively (
Table 8). The better performances of the COSVI-RPPs were attributed to the stronger penetration ability of SAR. The good consistency between the predicted values and measured values was due to the facts that OSVIs can provide an accurate interpretation of crop LAI and biomass and RPPs are more sensitive to crop canopy structure. Both of these factors contributed to the improved estimations of LAI and biomass. The results indicated that the advantages of optical and radar data were integrated and then could be used to enhance their application value. It had great significance to promote the development and integration of optical and radar technology. The results revealed that EVI × RVI and MTVI2 × RVI could be used for robust estimates of LAI and biomass in winter wheat, and the other combined indices were also valuable (
Table 8). The result of Capodici
et al. [
27] was also confirmed by our study. In this study, the COSVI-RPPs were acquired according to the spectral reflectance and SAR backscattering mechanism information. These new combined indices were used to estimate canopy structural information (LAI and biomass). The new combined indices were better than the OSVIs and RPPs alone, but more investigations and validations are needed before their regional applications. In this study, the acquisition time had affected the HJ and radar data. The LAI and biomass changed little during the tillering (4 March and 5 March) and filling stages (16 May and 20 May). The difference in the HJ and radar data acquisition time was ignored. However, because LAI and biomass changed at the jointing and anthesis stages, the difference in the HJ and radar data acquisition time may have led to some estimation errors. In particular, winter wheat grows quickly during the jointing stage. The data acquisition time of HJ on April 7th and of the radar data on March 29th resulted in a reduction in the estimation accuracy of LAI and biomass. Therefore, we think that growth stages impacted the predictive power of the indices, as the plants show very different optical-chemical and structural properties at different growth stages. These differences should influence the optical and SAR signals during different growth stages. Therefore, the small difference in the HJ and radar data acquisition time should be considered to better estimate LAI and biomass in future research. The establishment of regression equations and experimental field data are necessary, and some important factors should be carefully considered in the data analysis. For example, the incidence angle of SAR largely influenced the vegetation backscattering information [
27,
36] and SAR backscattering information was influenced by the amount of precipitation. These factors may result in some errors in the estimation of water-related crop or soil parameters.
The results of the MSR and PLSR methods show that the COSVI-RPPs were highly related to LAI and biomass of winter wheat (
Table 9). The estimation accuracy of LAI and biomass was higher with the PLSR than the MSR method. Previous studies have shown that PLSR outperforms MSR in estimating biophysical and biochemical parameters [
11,
59]. This may be because the MSR method can be used to concentrate on some spectral band features with known links to the variables of interest [
59]. In comparison, the PLSR method fully considered the relationships between the covariance of spectral band features and biophysical variables by applying data compression into regression factors. Therefore, the PLSR obtained the best estimation accuracy of LAI and biomass. However, the MSR also has merit, particularly when taking into consideration the simplicity of its application. But the MSR and PLSR regression models are quite unstable when they are applied to the larger region even though the calibration results look good in our study. Therefore, the performance of MSR and PLSR regression models have to be carefully verified by using sufficient independent datasets of different crops and ecological regions, as this study was limited to winter wheat in the Yangling district of Shaanxi, China.
5. Conclusion
In this study, the optical spectral vegetation indices (OSVIs), radar polarimetric parameters (RPPs), combined optical spectral vegetation indices and radar polarimetric parameters (the product of optical spectral vegetation indices and radar polarimetric parameters (COSVI-RPPs)), and multiple stepwise regression (MSR) and partial least squares regression (PLSR) methods were investigated to determine the most accurate empirical regression equations for LAI and biomass estimation in winter wheat. The results of this study revealed the following conclusions. Strong relationships existed between LAI, biomass and OSVIs, RPPs, and the OSVIs and RPPs could be used to estimate LAI and biomass in winter wheat based on the relevant regression equations. We found a highly significant correlation between the new COSVI-RPPs (RVI × OSVIs and DERD × OSVIs) and LAI and biomass. The estimation accuracy of LAI and biomass was better using RVI × OSVIs and DERD × OSVIs than using the OSVIs and RPPs values alone. The MSR and PLSR methods were used to estimate LAI and biomass in winter wheat based on the results of the COSVI-RPPs. The results demonstrated that the PLSR regression equations based on the COSVI-RPPs resulted in a better estimation of winter wheat LAI (R2 = 0.76, RMSE = 0.61, and nRMSE = 18.13%) and biomass (R2 = 0.85, RMSE = 137.21 g/m2, and nRMSE = 18.96%). The MSR regression equations based on the COSVI-RPPs also resulted in good estimations of LAI and biomass of winter wheat. The LAI (R2 = 0.78, RMSE = 0.58, and nRMSE = 17.42%) and biomass (R2 = 0.87, RMSE = 134.68 g/m2, and nRMSE = 18.61%) model obtained the best estimation results based on all the COSVI-RPPs, OSVIs, and RPPs using PLSR.