1. Introduction
Crop residues, such as stalks, stems, leaves, and seed pods, are materials left on the surface of agricultural fields after harvest. Crop residue cover (CRC), as an indicator of the amount of crop residue left on the soil surface, plays an important role in evaluating tillage intensity [
1,
2]. CRC is a key input parameter in models to predict the impact of agricultural systems on soil organic carbon, greenhouse gas emissions, and crop production [
3,
4]. Therefore, it is critical to estimate CRC accurately, thereby enabling the evaluation of the effectiveness of tillage practice and promoting agricultural sustainability.
Currently, the measurement of CRC depends on manual survey-based methods, such as the line-point transect. These methods are time-consuming and labor-intensive, thus making systematic and continuous quantification of CRC over large areas difficult [
5,
6]. Alternatively, remote sensing is an efficient technique to acquire CRC spatially and temporally in a rapid, accurate, and objective manner [
7,
8]. Remote sensing techniques used to estimate CRC can be classified into empirical regression based on crop residue indices (CRIs), classification [
9,
10], spectral unmixing [
11], and spectral angle methods [
12,
13]. The most widely used among these methods is empirical regression constructed from a linear or nonlinear relationship between CRC and the CRIs, also known as “the CRI technique”.
A series of CRIs have been designed to improve the detection of CRC. One such CRI, cellulose absorption index (CAI), was developed by Nagler et al. [
14] based on a unique absorption feature associated with cellulose and lignin at 2100 nm in the spectra of crop residue. The CAI has proven very accurate in quantifying CRC [
15]. However, its application is limited because hyperspectral imagery is currently only available on airborne or proximal platforms. Various broadband indices have been developed for mapping CRC using multispectral satellite imagery, such as the Advanced Spaceborn Thermal Emission and Reflectance Radiometer (ASTER) and the Landsat family. The well-known ASTER CRIs include the lignin cellulose absorption index (LCA), and the shortwave infrared (SWIR) normalized difference residue index (SINDRI) calculated from SWIR bands [
16,
17], which are effective at measuring CRC due to the precise measurement of cellulose and lignin absorption features [
18]. The CRIs developed from the Landsat imageries are also primarily dependent on SWIR bands, such as normalized difference tillage index (NDTI) [
19], normalized difference indices (NDI, NDI5, NDI7) [
20], normalized difference senescent vegetation index (NDSVI) [
21], and normalized difference residue index (NDRI) [
22]. Previous studies showed that NDTI performed the best in most cases [
15], and NDRI showed an advantage in minimizing the effects of green vegetation [
22]. Most of the above indices are limited to normalized difference ratios of two bands. Index formulations, e.g., simple ratio and normalized difference ratio, and band selection have impacts on the performance of the estimation [
23]. It is likely that indices based on more than two bands and different formulations would lead to further improvements in the retrieval of variables from remote sensing imagery.
The Sentinel-2 satellites launched in 2015 and 2017 carry Multi-spectral Instruments (MSI) with 13 spectral bands. Apart from the three red-edge bands, the MSI bands have corresponding spectrum regions to the Landsat 8 Operational Land Imager (OLI). Moreover, Sentinel-2 data possess finer spatial and temporal resolutions, which is more suitable for agricultural monitoring. However, there is a lack of studies on evaluating the ability of Sentinel-2 data in CRC estimation [
4,
24].
Besides the empirical regression approach, various machine learning (ML) regression algorithms have been developed and are popular in bio-geophysical variable retrieval due to their computational efficiency and effectiveness [
25,
26]. ML algorithms have shown advantages in capturing the non-linear relationship of input features and retrieval targets and can be massively multivariate, involving several variables. Commonly used ML methods include artificial neural network (ANN), support vector machine (SVM), Gaussian processes regression (GPR) [
27,
28,
29]. To date, very few studies have investigated the use of ML regression for the prediction of CRC [
30]. Therefore, evaluating the performance of ML regressions in CRC estimation is urgently needed. CRIs that are designed to enhance the discrimination of crop residue from soil are reportedly more correlated to CRC than spectral reflectances [
31]. Therefore, it is worth exploring the use of CRIs as input features in the ML models for the prediction of CRC. Furthermore, several studies reported that textural features were useful in estimating vegetation parameters (i.e., leaf area index (LAI)) and combining both textural features and spectral information provided much improvement compared with using spectral information only [
32,
33,
34]. In addition, Jin et al. [
35] evaluated three processing techniques to estimate CRC: regressions based on CRIs, textural features, and combinations of CRIs with textural analyses. They found that textural features were correlated with CRC and the regression based on a combination of textural features and CRIs yields a better result than the other two approaches. Based on these studies, it is necessary to evaluate the potential of textural features used as input predictors in the machine learners for the prediction of CRC.
To address the gaps, this study focused on evaluating the potential of Sentinel-2 data on CRC estimation using an empirical regression approach and ML regression techniques. We hypothesize that using different combinations of MSI bands, CRIs, and textural features could provide improved estimates of CRC than using any single source feature alone. Therefore, we evaluated and compared the performance of univariate regression against individual MSI bands, CRIs and texture features, as well as the retrieval accuracies of partial least squares regression (PLSR), ANN, support vector regression (SVR), GPR, and random forest (RF) associated with MSI bands, CRIs, and their combinations with textural features.
4. Discussion
In this study, we investigated empirical regression and ML methods for the purpose of estimating CRC from Sentinel-2 MSI data. Regression analysis using established CRIs and proposed CRIs based on Sentinel-2 MSI data were explored, together with the applications of PLSR and four non-linear non-parametric ML approaches (ANN, SVR, GPR, and RF), using (1) nine MSI bands, (2) 13 CRIs, (3) nine bands and four mean textural features, and (4) 13 CRIs and those textural features as predictor variables, respectively. Overall, the results demonstrated that PLSR, SVR, GPR, and RF provided higher retrieval accuracy when textural features and CRIs were used as input predictors. The GPR model performance was further improved by using optimized input predictors. GPR is recognized as a promising method for estimating CRC. A further discussion of the elements of this study is presented below.
4.1. Empirical Regression Analysis for CRC Estimation
Univariate regressions examining the relationship between individual bands, CRIs, and mean textural features and CRC were conducted. Most of the bands, CRIs, and textural features were correlated to CRC. The best-performing index among the published CRIs was NDRI with an accuracy of
R2cv of0.61 and RMSE
cv of 6.663%. The proposed index 3BI2, constructed with band 2, 4, and 12, obtained the best performance with
R2cv = 0.63 and RMSE
cv = 6.509%, followed by 3BI1 and 3BI3. These three-band indices performed better than two-band indices listed in
Table 1. Our results confirmed that, as noted by Verrelst et al. [
23], when multiple bands are available there is no reason to limit the estimation to two-band indices. Notwithstanding the promising results, we acknowledge that further comprehensive work is needed to evaluate the reliability of the three proposed indices to estimate CRC in the context of different residue types under various circumstances, such as residue decomposition stage, residue and soil moisture content, soil brightness, and the presence of green vegetation [
7,
18].
4.2. Machine Learning Approaches for CRC Estimation
In this study, five ML techniques (PLSR, ANN, SVR, GPR, and RF) were evaluated and the importance of input variables for these approaches was investigated. Moreover, the possibility of optimizing a GPR model for improving the retrieval accuracy of CRC was explored.
4.2.1. Impact of Training Samples on ML Algorithm Performance
ML algorithms typically require a large sample set to build an effective model. In order to analyze the effect of the number training samples on ML performance, each ML’s ability to handle limited training samples was examined. The same training samples were used for each algorithm to compare ML performance against a matching set of variables. A uniform percentage of the total samples for each algorithm was chosen at twenty percent intervals starting at 40% and ending at 80%. The sampling for each set was repeated 50 times. The models were trained with inputs of nine bands and four mean textures as an example. The results are presented in
Table 7. Results demonstrate that with bands and texture measures as input variables, the best-performing model is SVR across all training sample size trials. The increase in training sample size improved the accuracy for CRC estimation, starting at
equal 6.544% for 40% training samples of the whole sample population and improving to 6.354% for an 80% training sample size. The
increased with the increase in training samples. In general, all ML approaches improved accuracies with the increase in training sample size relative to total sample population.
4.2.2. Contribution of Input Predictors on ML Accuracy
First, the five methods were trained and run with MSI bands as input features, yielding significantly higher accuracies than regressions based on individual bands. When the 13 CRIs from
Table 1 were used as input features for a comparative analysis, the performances of the five models were further improved. However, regardless of whether nind MSI bands or 13 CRIs were used as input predictors for PLSR, ANN, GPR, and SVR, none of these models performed better than the univariate regression based on 3BI2 in terms of RMSE. When using mean textures and bands as input variable, the GPR, SVR, and RF models achieved better results than 3BI2. These three models also achieved a good performance by using mean textures and CRIs as predictors. Among the limited studies on the use of remote sensing indices as input features in ML models for the prediction of vegetation variables, Shah et al. [
28] showed the possibility of using vegetation indices in the RF for estimating chlorophyll. Our results showed the potential for using CRIs and textural features together as input variables in ML approaches for the enhanced estimation of CRC. Moreover, these results indicated that input predictors have an impact on the performance of ML models and highlighted the importance of mean textural feature to obtain better estimates of CRC.
The PLSR is a powerful tool that can model several response variables simultaneously while effectively addressing strong collinear variables [
52]. The retrieval accuracy of PLSR was improved by combining mean textures with bands. A combination of mean textures and CRIs via PLSR also enhanced the estimation of CRC. We extracted mean textural features from Sentinel-2 images, and correlated then with CRC. The results demonstrated that a combination of CRIs and mean textures via PLSR can enhance the retrieval accuracy of CRC.
ANN underperformed the other four ML models in terms of RMSE, regardless of the input features. This is likely to do with the fact that ANN is often used with large training datasets [
53] and it does not perform well when used with input variables deviating from the small dataset presented during the training stage [
54,
55]. In this study, we had a limited size of training dataset (
n = 243).
Recently, SVR gained popularity for the estimation of vegetation biophysical variables. Yang et al. [
56] applied the SVR to estimate leaf area index and leaf chlorophyll density of rice, and found that SVR outperformed linear non-parametric methods. Our result showed that when bands or CRIs were used as input variables, SVR did not perform as well as GPR and RF. However, the SVR model based on a combination of bands and textures outperformed the other methods.
Among the four non-linear non-parametric techniques, GPR and RF have the ability to optimize input variables for the purpose of enhancing model simplicity and improved accuracy. We did not investigate the importance of input variables for RF model because many studies have already done so [
28,
57]. We emphasized the application of the σ in the GPR covariance function to optimize input predictors. Our results demonstrate that the use of the optimal input variables filtered by σ for each groups of variables enhanced the prediction accuracy of GPR. From
Figure 7a, the MSI bands identified as highly important in the estimation of CRC were B3, B4, B5, B6, B11, and B12. When the nine MSI bands and four textural features were combined (
Figure 7c), B12 was also identified as the optimal variable. B12 (2190 nm) is near 2100 nm, where the reflectance of crop residue was mainly controlled by lignin and cellulose [
58]. Moreover, indices in
Table 1 associated with B12 provided a good retrieval of CRC (
Table 3). It is likely that the selection of B12 can provide useful information about CRC. From
Figure 7b, NDI5, NDRI, SGNDI, NDSVI, and 3BI1 were identified as the optimal input variables. It is worth noting that some of the CRIs that had high correlations with CRC were not identified as final input variables, such as 3BI2 and 3BI3, according to the values of σ. However, 3BI2 was selected as one of the optimal input variables when CRIs and textural features were combined (
Figure 7d). Such a result agrees with the conclusion by Shah et al. [
28] that remote sensing indices may perform variously when used in combination as input variables in a machine learner. Three of the four textural features were identified as optimal input variables not only when used in combination with bands but also when used in combination with CRIs (
Figure 7c,d). The best input features for GPR were B4, B8a, B12, and mean textures of B3, B4, and B8. The results highlight the power of the GPR algorithms for identifying the optimal information for CRC estimation and the importance of selecting proper input variables for obtaining robust outputs.
4.3. Importance of Textural Features for CRC Estimation
Image texture is a quantification of the spatial variation of image tone values, which can be related to spatial distribution of vegetation [
59]. The above-ground organization of crop residue elements is represented in texture, which is supplementary to the spectral image and may provide additional information about crop residue. In this study, although mean textural features extracted from Sentinel-2 10 bands have produced a low correlation with the UAV-derived CRC, we demonstrate that textural features have a high capability to provide an improvement in estimating CRC along with spectral information such as spectral bands and CRIs. For each ML algorithm, the combination of mean texture and MSI bands resulted in improvement in estimating CRC compared with MSI bands only. The combination of mean texture and CRIs also led to increasing estimation accuracy. Our experimental result confirmed Jin et al.’s findings [
35]. They compared the power using eight textural measures extracted from Landsat 8 OLI bands to map CRC with CRIs based on PLSR. Their study showed that the combination of textural features and CRIs performed better than CRI-based variables for estimating CRC. Such similar conclusions suggest that textural information along with spectral information could potentially improve estimating CRC compared to using spectral information only. The benefits of combining textural measures and spectral information in estimating vegetation canopy parameters (i.e., LAI and vegetation fractional coverage) have been proved [
32,
33,
34]. Zhou et al. [
32] evaluated the performance of vegetation indices (VIs), texture measures, and combinations of VIs and texture measures on LAI estimation. They found that the approach based on a combination of VIs and textural feature yielded higher accuracy than the VI-based approach and texture-based approach. Based on these studies, it can be concluded that the accuracy of estimated variables based on remote-sensing data could be increased by considering textural information.
4.4. Limitations of the Experiment
We presented a low-cost method for mapping CRC at ultrahigh resolution in this study. UAV-derived RGB orthomosaics were classified into two classes: crop residue and background. Our results showed a high classification accuracy, which provides a good basis for training and evaluating Sentinel-2 data-based models. It should be noted that the inversion only relied on UAV-CRC data. Several studies demonstrated that the UAV orthomosaics interpretation enables the cost-effective creation of large and comprehensive datasets in comparison to field work [
60,
61]; therefore, field measures of CRC were not included in this paper, given the ultrahigh resolution of UAV data. We acknowledge that there is space for improvement in the classification accuracy by comparing different classification methods [
62], such as object-based image analysis, or logistic regression, which is a powerful statistical learning method for a two-class situation [
63].
5. Conclusions
In this study, we focused on the estimation of CRC based on empirical regressions and machine learning methods from Sentinel-2 imagery. Based on UAV collected CRC and near-synchronous Sentinel-2 image, inversion models were established and validated. The results show that 3BI1, 3BI2, and 3BI3 improved the sensitivity to CRC and the estimation accuracy compared to the published CRIs. MSI bands, CRIs, and their combinations with mean textural features were used as input variables for PLSR, ANN, GPR, SVR and RF models. The five ML approaches with bands as input variables yielded accuracies second to those methods based on CRIs, the combination of indices and textural features, and the combination of bands and textural features, respectively. The SVR with bands and textural features yielded excellent performance (R2cv = 0.67, RMSEcv = 6.343%). The GPR based on CRIs and textural features also obtained high accuracy (R2cv = 0.66, RMSEcv = 6.352%). The retrieval accuracy of each GPR model optimized by was further improved. The best-performing model was a GPR with input variables of red, NIR, the second SWIR bands, and mean textural features of blue, red, and NIR bands, obtaining an accuracy of R2cv equal 0.69 and RMSEcv equal 6.149%. Comparing empirical regression and machine learning methods, it can be concluded that (1) most of the five ML models with the optimal input variables performed better than univariate regression, (2) using a scheme of combining mean textural features with spectral information could lead to higher accuracy of estimating CRC than spectral information alone, and (3) GPR is recommended for CRC estimation.