1. Introduction
Crop residue is defined as stalks, leaves and any other plant litter of crops such as maize, wheat, soybean, rice, etc., which accumulates on the surface of farmland after harvest [
1,
2]. Crop residues play an important role in the conservation of tillage systems, which include reducing water-based and wind-based soil erosion, increasing soil organic matter content, fixing CO
2 in the soil and improving soil quality [
3,
4,
5]. It also has positive influences on water infiltration and evaporation by improving soil structure and reducing daily variation of soil temperature [
6,
7]. Consequently, good residue management practices contribute significantly to an increase of crop yields [
8,
9,
10]. From the view point of air quality protection, remaining crop residues in farmland can greatly reduce the emission of toxic gasses such as CO, SO
2 and NH
3, while burning straws severely pollute the air [
11,
12]. Moreover, crop models need crop residue coverage (CRC) as an input parameter to simulate the impact of management practices on crop production. Therefore, it is of great importance to estimate CRC in crop planting areas and have a knowledge of the population of conservation tillage on a regional scale.
Remote sensing is a useful tool to estimate CRC efficiently on a regional scale in a timely manner [
13,
14,
15,
16]. Many attempts have been done to explore the relationship between CRC and satellite derived variables. Among them, regression-based spectral models were the most commonly used. They are based on the crop residue indices derived from optical remote sensing images, which included Normalize Difference Tillage Index (NDTI) [
17], Normalized Difference Senescent Vegetation Index (NDSVI) [
18], Normalized Difference Residue Index (NDRI) [
19], Normalized Difference Index 5 (NDI5) [
20], Normalized Difference Index 7 (NDI7) [
20], and so on. Previous studies have proved that NDTI performed best in most cases [
21], while NDSVI was more suitable in humid surroundings which avoided using the infrared bands (2080–2350nm) of Landsat TM. In the meanwhile, NDRI was proposed for the situation where green vegetables have appeared in the field. Thus, crop residue indices based on multispectral images are an integral part of the CRC estimation. However, as other indices used in the field of forest biomass estimation, and other quantitative remote sensing, crop residue indices have a poor performance in high coverage areas, which was named as the "saturation" phenomenon [
22]. It largely underestimated CRC in the range of 80–100% and influenced the overall accuracy of CRC estimation [
16,
22].
Compared with optical remote sensing information, microwave remote sensing has the advantage of not being sensitive to weather conditions. Previous studies have made good progress in CRC estimation based on microwave remote sensing information [
23,
24,
25,
26,
27]. The ground experiments have found that the backscattering coefficient of microwave scatterometer in a specific polarization direction has a good correlation with CRC of a specific crop type, which made it possible to estimate CRC by Synthetic Aperture Radar (SAR) data [
23,
24]. McNairn et al. [
25] also demonstrated the effectiveness of spaceborne SAR data in CRC estimation by using Spaceborne Imaging Radar-C data, and the parameters of cross-polarized linear backscatter and co-polarized circular backscatter were proved to be suitable for CRC estimation. Therefore, SAR imagery has great potential to estimate CRC. However, McNairn et al. also highlighted that the contribution of crop residues to radar response was only 40% [
25], which made it difficult to estimate crop residue cover using microwave responses alone. The estimation accuracy was easily influenced by residue type and condition, incidence angles, surface roughness and so on [
24,
27,
28]. All of these problems limited the use of SAR image in CRC estimation.
The estimation of CRC using optical or microwave data alone often has their own shortcomings. Studies have shown that the method combining optical and microwave remote sensing information had a better result than methods only using one type of remote sensing data in the fields of forest biomass estimation, leaf area index (LAI) estimation and other fields [
29,
30,
31,
32,
33,
34,
35]. Jin et al. [
29] estimated LAI and biomass of winter wheat with combined optical spectral vegetation indices and radar polarimetric parameters, which provided a guideline for the estimations of LAI and biomass using multisource remote sensing data. Huang et al. [
30] proved that the synergistic use of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and European Remote-Sensing Satellite-2 (ERS-2) SAR could estimate above ground biomass more accurately than using one type of remote sensing information in Xixi National Wetland Park, China. Cutler et al. [
31] combined the textural information of JERS-1 SAR image with Landsat TM to estimate forest biomass in three regions and the estimation accuracy also had significantly improvements. Synergistic use of optical imagery and SAR imagery can significantly improve the predicted accuracy of LAI and biomass. Therefore, exploring the effectiveness of the combined method for CRC estimation is of crucial importance.
Sentinel-1 and Sentinel-2 were both launched by the European space agency from 2014 to 2017. They have opened up opportunities for techniques which are aimed at improving the accuracy of CRC estimation. Sentinel-1 is a C-band synthetic aperture radar mission that provides data covering most parts of the world. Compared with other radar satellites in orbit, Sentinel-1 has a higher spectral resolution of 1dB/3σ, which makes it more suitable for quantitative estimation, and the wavelength of Sentinel-1 is relatively short. That makes it more likely to be effective for tillage assessment [
28]. Sentinel-2 is a multispectral satellite mission that sets up 13 bands from coastal wavelengths to short-wave infrared wavelength with spatial resolutions of 10, 20, or 60 m. Sentinel-2 can provide much more spectral information than Landsat OLI, which has made it more advantageous in quantitative estimation. Both Sentinel-1 and Sentinel-2 missions are in constellation with two twin satellites. Hence, the time resolution of Sentinel satellites can reach up to 6 days. The balance between spatial resolution and time resolution make the Sentinel-1 and Sentinel-2 more suitable for agricultural monitoring. Therefore, using Sentinel-1 and Sentinel-2 to explore the effectiveness of the method combining optical and microwave data for CRC estimation may have more meaning in the future applications.
Winter wheat is a main crop in the North China Plains, which is one of the most important food production regions in China. Estimating winter wheat residue coverage is important for agricultural production and environment protection. Therefore, the objectives of this study were to (i) investigate the potential of Sentinel-2 and Sentinel-2 for CRC estimation using several optical crop residue indices (OCRIs) and five radar parameters (RPs), (ii) investigate the relationship of CRC with new indices which combined OCRIs with RPs, and (iii) build up an optimal model for CRC estimation using OCRI-RPs by optimal subset regression. The findings of this study were expected to provide a guideline for accurately estimating CRC based on optical information and SAR information on a landscape scale.
4. Discussion
In this study, most of the optical crop residue indices used had certain correlations with field measured CRC, particularly NDTI (
Table 3). NDTI had a highest correlation with field measured CRC (R
2 = 0.570 and RMSE = 6.560%). This result was coincidental with the study of Jin et al. in 2015 [
37]. Previous studies have shown that crop residues and soil had different reflection characteristics in the 1450–1960 nm and 2000–2100 nm ranges [
1,
44]. The reflectance curve of winter wheat crop residues and soil collected by an Analytical Spectral Device (ASD) spectrometer in the laboratory (
Figure 9) also proved these reflection characteristics for the reflectance of crop residues having a peek of 1650 nm, and a valley of 2100 nm, while that of soil were flat and had a peak. Daughtry et al. [
1] indicted that water absorption in the 1450 nm and 1960 nm range made crop residues have a peak of reflectance in 1650 nm. The broad absorption feature of crop residues in 2100nm may be associated with cellulose and lignin. NDTI were calculated from the bands b11 (1540–1680 nm) and b12 (2080–2320 nm) of Sentinel-2, which were sensitive bands for crop residue identification. Thus, NDTI performed better than any other OCRIs. NDRI was calculated by the b4 (645–683 nm) and b12 of Sentinel-2. It had a certain ability to estimate CRC. Gelder et al. [
19] proved that NDRI performed better than NDTI when green vegetables appeared in farmland. However, in this study, there were little green vegetables on the farmland. Thus, the use of b4 made it more difficult to estimate CRC. NDI7 was calculated by the b8 (763–907 nm) and b12 bands of Sentinel-2. It has been proven that near infrared bands were sensitive to plant structure [
37,
45], which gives NDI7 the potential to distinguish crop residues from soil. Among the four new OCRIs, NDI71 performed better than NDI7, while NDI72, NDI73 and NDI74 performed worse than NDI7. It suggested that narrow bands were more sensitive to the reflectance of winter wheat. Therefore, more attention should be paid to the use of narrow bands for CRC estimation.
This study has proven that backscattering coefficients in the VV polarization direction and VH polarization direction have certain correlations with field measured CRC (R
2 = 0.341, RMSE = 8.086% and R
2 = 0.319, RMSE = 8.241%). The results were similar to the findings of McNairn et al. in 1992, who held the view that cross-polarized backscattering coefficients can estimate CRC well and backscattering coefficients in the VV polarization direction can distinguish crop residues from soil [
24]. However, the R
2 values were not very high between CRC and backscattering coefficient. It may because the incidence angles of Sentinel-1 images used in this study were both around 39°, but the best angle for CRC estimation was between 40° and 50° [
26]. Radar indices have been popular indices for vegetable water content estimation and fresh weight estimation [
46,
47]. This study tentatively explored the potential of radar indices for CRC estimation, and interestingly found that RI1 and RI2 both performed better than
and
, which proved the effectiveness of radar indices in CRC estimation. Sentinel-1 images data only have two polarization combinations. It has limited the study on exploring the relationships between CRC and radar indices constructed by backscattering coefficients in the other two polarization direction, which needed to be further explored in other microwave data.
Based on the well performed OCRIs and RPs, there were twenty OCRI-RPs which have been developed to estimate CRC. Most of the OCRI-RPs had a relatively higher R2 value and a relatively lower RMSE value when compared with the OCRIs and RPs, which suggested that it was effective to improve the accuracy of CRC estimation by combining optical information and microwave information. For the purpose of further improvement of model accuracy, this study used optimal subset regression to estimate CRC. As an alternative method of stepwise regression technology, optimal subset regression can well fit the experimental data. It uses all combinations of independent variables to fit dependent variables, which can solve the problem of inconsistency made by forward stepwise regression and backward stepwise regression. The final results showed that the combinations of NDI71 × and NDTI × had the highest correlations with field measured CRC (R2 = 0.770, RMSE = 4.846%). It had significant improvements in CRC estimation, when compared with the best results with OCRIs (R2 = 0.570 and RMSE = 6.560%), the best results with RPs (R2 = 0.430 and RMSE = 7.052%) and the best results with OCRI-RPs (R2 = 0.738, RMSE = 5.140%). It was seen that synergistic use of optical and microwave data is very helpful to improve the accuracy of CRC estimation, which can open a new way for CRC estimation.
As the result shows in
Figure 8, CRC in Yucheng County has been divided into three classes: 0–60%, 60–70%, and 70–100%. It could be easily found that most of the farmland in Yucheng County had a CRC between 60% and 70%. In order to further explore the relationship between the results and various satellite derived variables, this study calculated the mean values of some satellite derived variables in three classes (
Table 6). The results showed that the mean values of NDTI, NDI71 ×
, NDTI ×
were positively correlated with CRC. The results of NDTI were coincidental with the results in
Figure 4. It proved that NDTI, NDI71 ×
, NDTI ×
were good variables to estimate CRC. The mean value of NDI71 between 60% and 70% and the mean value of NDI71 between 70% and 100% were similar to results in
Figure 4. However, it did not have an appropriate mean value between 0% and 60%, which means it performed poorly in the estimation of low coverage areas. The mean values of RI1 and RI2 did not have a good relationship with CRC. They also performed poorly in low coverage areas. In conclusion, NDTI was the best variables in CRC estimation when just using spectral information or SAR information.