3.2.2. Non-Parametric Modeling Based on the Continuum Removal of Red-Edge Spectra

To investigate the influence of bandwidth on the predictive performance of continuum removal of red-edge spectra (SpeCR), both PLSR and LSSVM models were established based on the SpeCR under different bandwidths. In the comparison of Figure 4a,b, it could be observed that the predictive performance of the LSSVM regression models was generally better than that of the PLSR regression models. The RMSE values for the validation dataset generally increased with increasing bandwidth (Figure 4a,b). When the bandwidth exceeded 11 nm, the increasing RMSE values indicated that both LSSVM and PLSR models had an obvious decrease in predictive performance. Interestingly, both PLSR and LSSVM analyses on the SpeCR with a bandwidth of 1 nm did not produce the highest estimation accuracy. The PLSR model based on the SpeCR yielded the highest accuracy when the bandwidth was 5 (R<sup>2</sup> val = 0.83, RMSEval = 133.45 g/m<sup>2</sup> , and MAEval = 101.65 g/m<sup>2</sup> ).

The LSSVM regression model based on the SpeCR achieved the highest accuracy when the bandwidth was 9 (R<sup>2</sup> val = 0.87, RMSEval = 123.98 g/m<sup>2</sup> and MAEval = 92.08 g/m<sup>2</sup> ). Compared with the linear model based on the optimal NDVI-like, the optimal models based on the SpeCR significantly improved the estimation accuracy and decreased the RMSE value by 63.85~73.32 g/m<sup>2</sup> . As observed in the scatter plots between observed and estimated AGB (Figure 4c,d), to some extent, the models based on the SpeCR alleviated the underestimation problem for the AGB exceeding 800 g/m<sup>2</sup> , especially the LSSVM regression model. The optimal band combination with the largest R2 was composed of 1193 and 1222 nm. The univariate linear regression model based on the optimal NDVI-like (1193, 1222 nm) indices was constructed and assessed using the validation dataset. The predictive performance of this model is shown in Figure 3b. The RMSE of this model in the validation dataset reached up to 197.30 g/m2. The majority error resulted from the underestimation for the biomass samples exceeding 800 g/m2. It indicated that the optimal NDVI-like (1193, 1222 nm) suffered from saturation problem and consequently resulted in reduced winter wheat AGB estimation accuracy. **Figure 3.** (**a**) Two-dimension scalogram illustrating the coefficient of determination (R2) between biomass and narrowband NDVI-like indices based on all possible two-band combinations; (**b**) predictive performance of the optimal NDVI-like (1193, 1222 nm) for winter wheat biomass estimation. 3.2.2. Non-Parametric Modeling Based on the Continuum Removal of Red-Edge Spectra To investigate the influence of bandwidth on the predictive performance of contin-

uum removal of red-edge spectra (SpeCR), both PLSR and LSSVM models were estab-

Narrowband NDVI-like indices were calculated using all possible two-band combinations. The univariate linear regression analyses of winter wheat biomass against these NDVI-like indices were performed to obtain the R2 values. As shown in Figure 3a, the formed two-dimension (2D) correlation scalogram exhibits a characteristic pattern with multiple "hot spots" with relatively high R2 values. The spots were selected by choosing the wavelength combination that had an R2 larger than 0.55 (*p* < 0.01). It could be observed that most of the wavebands forming these "hot spots" were located in the spectral range of 1168⁓1276 nm, and several band pairs were located in the spectral range of 675⁓683 nm.

*Remote Sens.* **2021**, *13*, 581 11 of 22

(Std.) of the observed AGB are also shown in Table 2.

**Table 2.** Descriptive statistics for winter wheat AGB (g/m2) of calibration and validation datasets. **Wheat AGB Min. Mean Max. Std. CV (%) Kurtosis**  Calibration 50.06 575.85 1759.95 333.30 57.88 3.38 Validation 75.96 526.17 1501.63 320.38 60.88 3.25

*3.2. Using Spectral Features to Estimate Winter Wheat AGB* 

*Remote Sens.* **2021**, *13*, 581 12 of 22

3.2.1. Univariate Linear Regression Based on Narrowband NDVI-Like

The differences in winter wheat aboveground biomass were created by multiple treatments including different varieties, different levels of N fertilizer applications, and different irrigation rates. The AGB values under different treatments were compared through the least significant different test (LSD) at the 95% level of significance. The results showed that there existed significant difference in the winter wheat AGB collected from the three experiments (*p* < 0.05). The descriptive statistics of winter wheat AGB exhibited a high degree of variation with a coefficient of variation (CV) of 57.88% for the calibration dataset and a CV of 60.88% for the validation dataset (Table 2). The range of AGB in the calibration dataset is larger than that for the validation dataset. The kurtosis values indicated that the AGB in both calibration and validation datasets approximately followed normal distribution. To help to understand the RMSE and MAE for calibration and validation, the minimum (Min.), mean, maximum (Max.), and standard deviation

**3. Results** 

*3.1. Descriptive Statistics* 

**Figure 3.** (**a**) Two-dimension scalogram illustrating the coefficient of determination (R<sup>2</sup> ) between biomass and narrowband NDVI-like indices based on all possible two-band combinations; (**b**) predictive performance of the optimal NDVI-like (1193, 1222 nm) for winter wheat biomass estimation. RMSE value by 63.85⁓73.32 g/m2. As observed in the scatter plots between observed and estimated AGB (Figure 4c,d), to some extent, the models based on the SpeCR alleviated the underestimation problem for the AGB exceeding 800 g/m2, especially the LSSVM re-

gression model.

**Figure 4.** Predictive performance of the continuum removal of red-edge spectra (SpeCR) for winter wheat biomass estimation. Change of the predictive performance of SpeCR using PLSR (**a**) and LSSVM (**b**) under different bandwidths. Predictive performance of the SpeCR using PLSR at the optimal bandwidth (FWHM = 5) (**c**) and using LSSVM at the optimal bandwith (FWHM = 9) (**d**). **Figure 4.** Predictive performance of the continuum removal of red-edge spectra (SpeCR) for winter wheat biomass estimation. Change of the predictive performance of SpeCR using PLSR (**a**) and LSSVM (**b**) under different bandwidths. Predictive performance of the SpeCR using PLSR at the optimal bandwidth (FWHM = 5) (**c**) and using LSSVM at the optimal bandwith (FWHM = 9) (**d**).

#### *3.3. Using Multiscale Textures to Estimate Winter Wheat AGB 3.3. Using Multiscale Textures to Estimate Winter Wheat AGB*

The Pearson correlation analyses were conducted between textures and winter wheat AGB during multiple growth stages (Figure 5). For the single-scale GLCM-based textures, the texures from blue band showed higher correlation with AGB than the textures from red and green bands. The texture Var-B produced the highest correlation coefficient (r = 0.41, *p* < 0.01). For the textures extracted by the proposed Multiscale\_Gabor\_GLCM method, the textures correlating well with AGB could be observed at each scale (Figure 5b). The texture Mea-G-S1 showed the strongest correlation with AGB (r = 0.69, *p* < 0.01). As can be observed from the scatter plots shown in Figure 5c,d, there exist linear relationships between these best-performing textures and AGB for each growth stage. However, the sensitivity of these textures to the variation in AGB decreased when considering multiple growth stages. In the comparison of Figure 5c,d, the best-performing texture Mea-G-S1 exhibited higher correlation with AGB than did the best-performing GLCM-based texture Var-B. For the AGB exceeding 800 g/m<sup>2</sup> , the texture Mea-G-S1 still stayed sensitive to AGB, while the texture Var-B approximately approached a saturation level and disorderly response to AGB. Therefore, the models based on the multiscale textures might be more promising than those based on the single-scale GLCM-based textures. The Pearson correlation analyses were conducted between textures and winter wheat AGB during multiple growth stages (Figure 5). For the single-scale GLCM-based textures, the texures from blue band showed higher correlation with AGB than the textures from red and green bands. The texture Var-B produced the highest correlation coefficient (r = 0.41, *p* < 0.01). For the textures extracted by the proposed Multiscale\_Gabor\_GLCM method, the textures correlating well with AGB could be observed at each scale (Figure 5b). The texture Mea-G-S1 showed the strongest correlation with AGB (r = 0.69, *p* < 0.01). As can be observed from the scatter plots shown in Figure 5c,d, there exist linear relationships between these best-performing textures and AGB for each growth stage. However, the sensitivity of these textures to the variation in AGB decreased when considering multiple growth stages. In the comparison of Figure 5c,d, the best-performing texture Mea-G-S1 exhibited higher correlation with AGB than did the best-performing GLCM-based texture Var-B. For the AGB exceeding 800 g/m2, the texture Mea-G-S1 still stayed sensitive to AGB, while the texture Var-B approximately approached a saturation level and disorderly response to AGB. Therefore, the models based on the multiscale textures might be more promising than those based on the single-scale GLCM-based textures.

**Figure 5.** Responses of different textures to winter wheat AGB in the calibration dataset. (**a**) Pearson correlation analyses between GLCM-based textures and AGB; (**b**) Pearson correlation analyses between AGB and Multiscale\_Gabor\_GLCM textures; (**c**,**d**) Scatter plots between the best-performing textures and AGB for each growth stage. **Figure 5.** Responses of different textures to winter wheat AGB in the calibration dataset. (**a**) Pearson correlation analyses between GLCM-based textures and AGB; (**b**) Pearson correlation analyses between AGB and Multiscale\_Gabor\_GLCM textures; (**c**,**d**) Scatter plots between the best-performing textures and AGB for each growth stage.

PLSR analyses were conducted uisng the single-scale GLCM-based textures and multiscale textures, respectively. Figure 6 shows the results of these models for both calibration and validation datasets. For the two PLSR models, the optimal latent factors were 11 and 10, respectively. Compared with the PLSR model based on the single-scale GLCM-based textures, the PLSR model based on the multiscale textures performed better with a R2 value of 0.78 , a RMSE value of 154.54 g/m2, and a MAE value of 122.86 g/m2. This model increased the R2 value by 0.09 and decreased the RMSE value by 24.58 g/m2. PLSR analyses were conducted uisng the single-scale GLCM-based textures and multiscale textures, respectively. Figure 6 shows the results of these models for both calibration and validation datasets. For the two PLSR models, the optimal latent factors were 11 and 10, respectively. Compared with the PLSR model based on the single-scale GLCM-based textures, the PLSR model based on the multiscale textures performed better with a R<sup>2</sup> value of 0.78, a RMSE value of 154.54 g/m<sup>2</sup> , and a MAE value of 122.86 g/m<sup>2</sup> . This model increased the R<sup>2</sup> value by 0.09 and decreased the RMSE value by 24.58 g/m<sup>2</sup> .

**Figure 6.** Predictive performance of the PLSR models using (**a**) the single-scale GLCM-based textures and (**b**) multiscale textures for winter wheat AGB estimation. **Figure 6.** Predictive performance of the PLSR models using (**a**) the single-scale GLCM-based textures and (**b**) multiscale textures for winter wheat AGB estimation.

The influential textures were determined during the construction of LSSVM regression models. According to the importance value of each texture (Figure 7a for the GLCM-based textures; Figure 7b for the multiscale textures), the GLCM-based textures with importance values larger than 5 and the multiscale textures with importance values larger than 200 were selected for the rebuilding of new LSSVM regression models after several trials. A total of seven single-scale GLCM-based textures were selected including Mea-R, Var-R, Con-R, Mea-G, Mea-B, Var-B, and Con-B. The selected multiscale textures were Mea-R-S1, Mea-R-S3, Mea-R-S4, Mea-R-S5, Mea-G-S1, Mea-G-S2, Mea-G-S3, Mea-G-S4, Mea-G-S5, and Mea-B-S1 (10 textures). It can be observed that the multiscale texture metric "Mea" played an important role in AGB estimation. Compared with the initial LSSVM regression models based on all textures, the new LSSVM regression models not only improved the predictive performance, but also reduced the complexity of models (See Figure A1 in Appendix A). The LSSVM regression model based on the selected multiscale textures exhibited higher accuracy with R2val = 0.85, RMSEval = 129.45 g/m2, and MAEval = 97.11 g/m2, compared to the LSSVM regression model using the selected singlescale GLCM-based textures (Figure 7c,d). For both single-scale GLCM-based textures and multiscale textures, the LSSVM regression models performed better than the PLSR models (Figures 6 and 7). For instance, compared with the PLSR model based on the multiscale textures, the LSSVM regression model using the selected multiscale features increased the R2 value by 0.08 and reduced RMSE value by 25.09 g/m2. It indicated that LSSVM regression was superior to PLSR in characterizing the nonliner relationship between textures and winter wheat AGB. Compared with the spectral features based models, the LSSVM regression model based on the multiscale textures performed better than the linear model based on the optimal NDVI, but just got comparable estimation accuracy with the LSSVM regression model based on the SpeCR with a bandwidth of 9. The influential textures were determined during the construction of LSSVM regression models. According to the importance value of each texture (Figure 7a for the GLCM-based textures; Figure 7b for the multiscale textures), the GLCM-based textures with importance values larger than 5 and the multiscale textures with importance values larger than 200 were selected for the rebuilding of new LSSVM regression models after several trials. Atotal of seven single-scale GLCM-based textures were selected including Mea-R, Var-R,Con-R, Mea-G, Mea-B, Var-B, and Con-B. The selected multiscale textures were Mea-R-S1, Mea-R-S3, Mea-R-S4, Mea-R-S5, Mea-G-S1, Mea-G-S2, Mea-G-S3, Mea-G-S4, Mea-G-S5, and Mea-B-S1 (10 textures). It can be observed that the multiscale texture metric "Mea" played an important role in AGB estimation. Compared with the initial LSSVM regression models based on all textures, the new LSSVM regression models not only improved the predictive performance, but also reduced the complexity of models (See Figure A1 in Appendix A). The LSSVM regression model based on the selected multiscale textures exhibited higher accuracy with R<sup>2</sup> val = 0.85, RMSEval = 129.45 g/m<sup>2</sup> , and MAEval = 97.11 g/m<sup>2</sup> , compared to the LSSVM regression model using the selected single-scale GLCM-based textures (Figure 7c,d). For both single-scale GLCM-based textures and multiscale textures, the LSSVM regression models performed better than the PLSR models (Figures 6 and 7). For instance, compared with the PLSR model based on the multiscale textures, the LSSVM regression model using the selected multiscale features increased the R<sup>2</sup> value by 0.08 and reduced RMSE value by 25.09 g/m<sup>2</sup> . It indicated that LSSVM regression was superior to PLSR in characterizing the nonliner relationship between textures and winter wheat AGB. Compared with the spectral features based models, the LSSVM regression model based on the multiscale textures performed better than the linear model based on the optimal NDVI, but just got comparable estimation accuracy with the LSSVM regression model based on the SpeCR with a bandwidth of 9.

To further demonstrate the necessity of extracting multiscale textures for AGB estimation under multiple growth stages, the LSSVM regression analyses were conducted using the textures at each scale to examine their predictive performance. According to the importance values for the textures at each scale, it could be observed that the texture metric "Mea" still had a higher degree of importance than the other texture metrics. Among the textures from five scales, the textures at scale 1 yielded the best estimation accuracy with R2val = 0.82, RMSEval = 140.95 g/m2, and MAEval = 109.10 g/m2. The textures at scale 2 and scale 3 got the intermediate estimation accuracy. The textures at scale 5 got the worst estimation accuracy and achieved comparable accuracy with the single-scale GLCM-based textures. It could be obviously observed that the LSSVM regression model based on the multiscale textures produced better estimation accuracy than that based on the single-scale textures. To further demonstrate the necessity of extracting multiscale textures for AGB estimation under multiple growth stages, the LSSVM regression analyses were conducted using the textures at each scale to examine their predictive performance. According to the importance values for the textures at each scale, it could be observed that the texture metric "Mea" still had a higher degree of importance than the other texture metrics. Among the textures from five scales, the textures at scale 1 yielded the best estimation accuracy with R 2 val = 0.82, RMSEval = 140.95 g/m<sup>2</sup> , and MAEval = 109.10 g/m<sup>2</sup> . The textures at scale 2 and scale 3 got the intermediate estimation accuracy. The textures at scale 5 got the worst estimation accuracy and achieved comparable accuracy with the single-scale GLCM-based textures. It could be obviously observed that the LSSVM regression model based on the multiscale textures produced better estimation accuracy than that based on the single-scale textures.

**Figure 7.** Predictive performance of the LSSVM regression models based on the selected GLCM-based textures (**a**,**c**) and selected multiscale textures (**b**,**d**) for winter wheat AGB estimation. **Figure 7.** Predictive performance of the LSSVM regression models based on the selected GLCM-based textures (**a**,**c**) and selected multiscale textures (**b**,**d**) for winter wheat AGB estimation.

#### *Estimation*  With combinations of selected multiscale textures, optimal NDVI and SpeCR, the *3.4. The Combined Use of Multiscale Textures and Spectral Features for Winter Wheat AGB Estimation*

*3.4. The Combined Use of Multiscale Textures and Spectral Features for Winter Wheat AGB* 

LSSVM regression analyses did not significantly improve estimation accuracy and produced generally better (or similar) results than (or to) those using textures or SpeCR alone. Among these models, the LSSVM regression model based on the combination of selected multiscale textures and SpeCR (FWHM = 9) yielded the highest estimation accuracy of AGB with R2val = 0.87, RMSEval = 119.76 g/m2, and MAEval = 91.61 g/m2 (Table 3). However, the combination of selected mulsticale textures and optimal NDVI just consisted of 11 features and yielded comparable estimation accuracy to the combination of multiscale textures and SpeCR (FWHM = 9) consisting of 29 features (i.e., 10 selected multiscale textures and 19 spectral features). With combinations of selected multiscale textures, optimal NDVI and SpeCR, the LSSVM regression analyses did not significantly improve estimation accuracy and produced generally better (or similar) results than (or to) those using textures or SpeCR alone. Among these models, the LSSVM regression model based on the combination of selected multiscale textures and SpeCR (FWHM = 9) yielded the highest estimation accuracy of AGB with R<sup>2</sup> val = 0.87, RMSEval = 119.76 g/m<sup>2</sup> , and MAEval = 91.61 g/m<sup>2</sup> (Table 3). However, the combination of selected mulsticale textures and optimal NDVI just consisted of 11 features and yielded comparable estimation accuracy to the combination of multiscale textures and SpeCR (FWHM = 9) consisting of 29 features (i.e., 10 selected multiscale textures and 19 spectral features).

**Table 3.** Performance of the LSSVM regression models based on the combination of multiscale textures and biomasssensitive spectral features for winter wheat biomass estimation. **Table 3.** Performance of the LSSVM regression models based on the combination of multiscale textures and biomass-sensitive spectral features for winter wheat biomass estimation.


#### **4. Discussion 4. Discussion**

A three-year field experiment with different cultivars, N fertilizer application, and irrigation rates led to a large number of samples with high variations in winter wheat AGB. An economical UAV remote sensing system successfully acquired the RGB digital A three-year field experiment with different cultivars, N fertilizer application, and irrigation rates led to a large number of samples with high variations in winter wheat

AGB. An economical UAV remote sensing system successfully acquired the RGB digital imagery of winter wheat canopy during growing seasons. The spatial resolution of the RGB imagery reached up to 1 cm. It allowed the extraction of fine textural features and the assessment of their predictive performance for winter wheat AGB estimation. The ground-based hyperspectral data and biomass measurements were concurrently collected. A comparison of the predictive performance of hyperspectral features, textural features, and combinations thereof for winter wheat AGB estimation was also possible.

#### *4.1. Performance Comparison of Simple and Complex Spectral Models for Wheat Winter AGB Estimation*

Using all the possible two-band combinations for narrowband NDVI-like calculation facilitated the selection of sensitive hyperspectral features for winter wheat AGB. In the linear regression analyses of narrowband NDVI-like indices against winter wheat AGB, it was found that the wavebands forming the "hot spots" in the 2D correlation matrix plot were mainly located between 1168 and 1276 nm (Figure 2). This spectral area is close to 1200 nm, which is affected by the absorption of water, cellulose, starch, and lignin [36]. Yao et al. [11] also indicated that the continuous wavelet features extracted from this spectral region exhibited strong correlation with wheat AGB. Gnyp et al. [37] developed a new hyperspectral vegetation index named GnyLi, which was calculated from wavebands 900, 955, 1050, and 1220 nm. They indicated that the GnyLi vegetation index correlated well with winter AGB during several growth stages (Z.S.30~45) and highlighted the ability of local peak spectra in near-infrared and shortwave infrared regions in alleviating the saturation problem for vegetation AGB estimation. In our study, the best-performing NDVI-like index was composed of wavebands 1193 and 1222 nm. However, in the study of Fu et al. [9], the best narrowband NDVI-like index for winter wheat AGB estimation was calculated from wavebands located at 980 and 1097 nm. The inconsistency in selected wavebands mainly ascribed to the use of different calibration datasets. It is a common problem in band selection for crop AGB estimation based on data-driven methods. In spite of this, the selected wavebands forming the optimal NDVIs can be interpretable in terms of spectral response of crop biomass. Further studies could focus on the improvement in the representativeness of calibration datasets. It can be realized by collecting wheat samples with different cultivars, growth stages, field treatments, and ecological conditions. Based on the comprehensive datasets, the selected sensitive wavebands might be more stable and the predictive model based these wavebands might have better generalization ability. Another way to improve the stability of narrowband vegetation indices is to develop the vegetation indices, which are specific to a certain crop cultivar and environment condition. Obviously, the latter ones are confined to limited application.

Both PLSR and LSSVM analyses using the continuum removal of hyperspectra between 550 and 750 nm produced better estimation accuracy as compared with the linear regression model based on the optimal NDVI (1193 nm, 1222 nm). The results confirm the capability of red-edge spectra in predicting winter wheat AGB, consistent with previous findings reported by Hansen and Schjoerring [3], Fu et al. [9], and Kanke et al. [38]. They have indicated that the red-edge region contains useful information relevant to crop biomass. It is important to note that the LSSVM regression analysis performed better than the PLSR for winter wheat AGB estimation, when combined with the continuum removal of spectra with different bandwidths. It demonstrated that LSSVM regression analysis was advantageous over PLSR in characterizing the relationships between spectral features and winter wheat AGB during multiple growth stages. In the exploration of the influence of bandwidth on the predictive performance of red-edge spectra, both the LSSVM regression and PLSR models generally got decreased estimation accuracy with increasing bandwidths. However, the continuum removal of spectra with a bandwidth of 1 nm did not yield the highest estimation accuracy. It mainly attributes to the high intercorrelation in neighboring hyperspectral wavebands, which has a negative impact on model accuracy and interpretability. However, PLSR and LSSVM have the ability to partly overcome the multi-collinearity problem. Through increasing the bandwidth, the spectral redundancy

can be reduced. The PLSR and LSSVM regression models achieved the highest estimation accuracy at the bandwidths of 5 and 9 nm, respectively. The obtained estimation accuracies were comparable to those of the PLSR and LSSVM regression models using all spectral wavebands between 400 and 1350 nm (Figure A2 in Appendix A). After the bandwidth exceeded 11 nm, the predictive performance of both PLSR and LSSVM models had a clear decrease (Figure 4a,b). It means that a fraction of spectral features associated with winter wheat AGB are lost when the hyperspectral data are resampled to multispectral data. It also demonstrates the necessity of exploring hyperspectral features for winter wheat AGB estimation. At the bandwidth of 21 nm, the predictive performance of the LSSVM regression model was still acceptable with a RMSE value of 132.05 g/m<sup>2</sup> . This result is conducive to develop hand-held multispectral devices or low-cost digital cameras for proximal crop monitoring.

#### *4.2. Performance Comparison of Single-Scale and Multiscale Textures for Winter Wheat AGB Estimation*

With the aim to extract textures that can well characterize the variations of winter wheat canopy during multiple growth stages, a multiscale texture extraction method was proposed (Multiscale\_Gabor\_GLCM). This method comprehensively utilized the advantages of Gabor filter and GLCM analyses. In the Pearson correlation analyses between winter wheat AGB and textures, the texture metric "Mea-G-S1" extracted by the proposed method exhibited higher correlation with AGB than the best-performing single-scale GLCM-based texture "Var-B" (Figure 5). Yue et al. [18] also reported that the GLCM-based texture metric "Var-B" correlated well with winter wheat AGB. Zheng et al. [17] indicated that the GLCM-based texture metric "Mea" was better than the other seven texture metrics in rice AGB estimation. Through the visualization technique of LSSVM regression, we determined 10 textures (i.e., metric "Mea" from different scales) that contributed most to winter wheat AGB (Figure 7b). In the exploration whether multiscale textures could show superiority to single-scale textures for winter wheat AGB estimation, both PLSR and LSSVM regression models based on the multiscale textures yielded higher R<sup>2</sup> and lower RMSE values for the validation dataset compared to the models based on the single-scale textures including GLCM-based textures and textures at each scale (Figures 6–8). The RGB imagery of winter wheat canopy is usually composed of soils, leaves, stems, ears, and shades resulted from leaf clumping. The proportion of each component changes and wheat canopy structure gradually becomes complicated during growing seasons. At the jointing stage, wheat canopy cannot fully cover soil background, and soils can be seen from imagery. From the heading stage to anthesis stage, wheat ears and flowers emerge, and wheat canopy almost fully covers soil background. From the anthesis stage to early grain filling stage, wheat ears become plump and leaves begin to turn yellow. These wheat canopy dynamic changes reflected in digital images can be captured by fine image textures. Moreover, winter canopies usually exhibit heterogeneity within or among fields due to different cultivars, field treatments, and cropland micro-meteorology, even at the same growth stage. Therefore, using single-scale textures will hardly capture the spatial changes of wheat canopies during growing seasons. This is confirmed by the predictive performance comparison between multiscale and single-scale textures for winter wheat AGB estimation. The proposed multiscale texture extraction method firstly utilized Gabor filters with multiple scales and orientations to generate textural images (i.e., Gabor magnitude images). These Gabor textures are less sensitive to variations of local lighting conditions, which will improve the reliability of subsequent image processing. Then GLCM analyses were conducted on these images to extract finer textures. Our previous study [20] has found that the textures of different orientations did not show significant heterogeneity for winter wheat canopy imagery; thus, this study only focused on the scale properties of textures derived from UAV-based RGB imagery. The optimal scale of textures is highly correlated with crop planting density and canopy size. However, the quantitative relationships among them are still unclear. To obtain abundant discriminative features, we followed the previous research results [20,32] and set five different scales,

which determined the spatial frequencies of Gabor filters. In the study of Yue et al. [18], they resampled original UAV-based RGB images to images with spatial resolutions of 2, 5, 10, 15, 20, 25, and 30 cm, respectively. Then they extracted GLCM-based textures from these images for winter wheat AGB estimation. The aim of doing this was also to extract multiscale textures to represent winter wheat canopy variations during multiple growth stages. They got improved estimation accuracy of winter wheat AGB by combining PLSR and the GLCM-based textures derived from RGB images with different spatial resolutions, which was consistent with the findings in our study. However, it seemed a little arbitrary in the determination of spatial resolution of images for analysis. In our proposed multiscale texture extraction method, the scale of textures was controlled by the scale parameter of Gabor filters. Further studies should investigate the quantitative relationships between crop canopy structure and scale parameters. It could help to determine the optimal scale parameter when extracting multiscale textures for crops of interest. In this study, the proposed multiscale texture extraction framework can also involve other multiresolution image analysis methods such as 2D wavelet transformation and 3D Gabor filter. termined the spatial frequencies of Gabor filters. In the study of Yue et al. [18], they resampled original UAV-based RGB images to images with spatial resolutions of 2, 5, 10, 15, 20, 25, and 30 cm, respectively. Then they extracted GLCM-based textures from these images for winter wheat AGB estimation. The aim of doing this was also to extract multiscale textures to represent winter wheat canopy variations during multiple growth stages. They got improved estimation accuracy of winter wheat AGB by combining PLSR and the GLCM-based textures derived from RGB images with different spatial resolutions, which was consistent with the findings in our study. However, it seemed a little arbitrary in the determination of spatial resolution of images for analysis. In our proposed multiscale texture extraction method, the scale of textures was controlled by the scale parameter of Gabor filters. Further studies should investigate the quantitative relationships between crop canopy structure and scale parameters. It could help to determine the optimal scale parameter when extracting multiscale textures for crops of interest. In this study, the proposed multiscale texture extraction framework can also involve other multiresolution image analysis methods such as 2D wavelet transformation and 3D Gabor filter.

we followed the previous research results [20,32] and set five different scales, which de-

*Remote Sens.* **2021**, *13*, 581 18 of 22

**Figure 8.** Importance values (**a**–**e**) and predictive performance (**f**–**j**) of the textures at each scale based on LSSVM regession analyses. **Figure 8.** Importance values (**a**–**e**) and predictive performance (**f**–**j**) of the textures at each scale based on LSSVM regession analyses.

#### *4.3. Evaluation of the Combined Use of Hyperspectral and Textural Features for Winter Wheat AGB Estimation*

Before the combined use of hyperspectral and textural features, the predictive performances of them were compared for winter wheat AGB estimation. Compared with the models based on the continuum removal of red-edge spectra (SpeCR, FWHM = 9), the models based on the multiscale textures had comparable estimation accuracy. It confirms the utility of textures derived from high-definition RGB images for winter wheat biomass estimation, since UAV-based RGB images are more readily acquired than ground-based, airborne, and spaceborne hyperspectral data. In the exploration of whether LSSVM analyses conducted using the combination of optimal NDVI, SpeCR (FWHM = 9), and multiscale textures could further improve estimation accuracy of winter wheat AGB, the LSSVM regression model based on the combination of the SpeCR (FWHM = 9) and selected multiscale features yielded the higher R<sup>2</sup> and lower RMSE for validation dataset, compared to the other combinations (i.e., multiscale textures + optimal NDVI and multiscale textures + optimal NDVI + SpeCR). However, the combination did not significantly improve the estimation accuracy compared to the use of SpeCR or multiscale textures only. The results are not in agreement with the findings reported by Yue et al. [18] and Zheng et al. [17], who highlighted that the combination of spectral and textural features could improve the estimation accuracy of crop AGB. The primary reason for this is that the multiscale textures in the study perform better than the GLCM-based textures used in [17,18]. The addition of red-edge hyperspectral information could not make an improvement in estimation accuracy. However, the spectral features used in [17,18] were broadband vegetation indices or some common narrowband vegetation indices, such as NDVI (670, 800 nm), optimized soil adjusted vegetation index (OSAVI (670, 800 nm)), and two-band enhanced vegetation index (EVI2), which did not utilize the informative information from red-edge spectra well.

#### **5. Conclusions**

This study investigated the utility of multiscale textures extracted from high-definition UAV-based RGB images in predicting winter wheat biomass during multiple growth stages. The proposed multiscale texture extraction method (Multiscale\_Gabor\_GLCM) took good advantage of Gabor filters and GLCM analyses. The feature selection function embedded in LSSVM regression helped identify 10 influential multiscale textures to winter wheat biomass. Both the PLSR and LSSVM regression models using the multiscale textures produced better estimation accuracies than those using single-scale GLCM-based textures. Compared to the spectral features based models, the multiscale textures got comparable estimation accuracy to the SpeCR with a bandwidth of 9 nm. Although the selected bestperforming NDVI (1193, 1222 nm) got intermediate estimation accuracy, it had great value in practical application due to its ease of use and low computation. The LSSVM regression analyses conducted using the combination of multiscale textures and hyperspectral features did not significantly improve the estimation accuracy for winter wheat AGB, compared to the use of multiscale textures or SpeCR alone. The LSSVM regression and its visualization technique showed great potential in establishing a high-accuracy winter wheat estimation model and understanding the driving force behind the model. This study confirmed the reliability of multiscale textures, red-edge spectra, and optimal NDVI calculated by watersensitive wavebands for winter wheat AGB estimation. Further studies are required to verify the effectiveness of multiscale texture extraction method in predicting crop biomass at different ecological regions, especially for the high-yielding crops.

**Author Contributions:** Conceptualization: Y.F.; methodology: Y.F.; software: Y.F.; validation: Y.F.; formal analysis: Y.F.; investigation: Y.F.; resources: G.Y. and H.F.; data curation: G.Y., X.X., H.F., and X.S.; writing—original draft preparation: Y.F.; writing—review and editing: Y.F. and G.Y.; visualization: Y.F.; supervision: G.Y., Z.L., and C.Z.; project administration: G.Y. and C.Z.; funding acquisition: Y.F, G.Y., and C.Z. All authors have read and agreed to the published version of the manuscript.

**Funding:** This study was supported by the Natural Science Foundation of China (41801225), the Key-Area Research and Development Program of Guangdong Province (2019B020214002, 2019B020216001), the National Key Research and Development Program of China (2017YFE0122500) and the Beijing Million Talent Project (No.2019A10). manuscript. **Funding:** This study was supported by the Natural Science Foundation of China (41801225), the Key-Area Research and Development Program of Guangdong Province (2019B020214002, 2019B020216001), the National Key Research and Development Program of China (2017YFE0122500) manuscript. **Funding:** This study was supported by the Natural Science Foundation of China (41801225), the Key-Area Research and Development Program of Guangdong Province (2019B020214002, 2019B020216001), the National Key Research and Development Program of China (2017YFE0122500)

acquisition: Y.F, G.Y., and C.Z. All authors have read and agreed to the published version of the

acquisition: Y.F, G.Y., and C.Z. All authors have read and agreed to the published version of the

**Institutional Review Board Statement:** Not applicable. and the Beijing Million Talent Project (No.2019A10). and the Beijing Million Talent Project (No.2019A10).

*Remote Sens.* **2021**, *13*, 581 20 of 22

*Remote Sens.* **2021**, *13*, 581 20 of 22

**Informed Consent Statement:** Not applicable. **Institutional Review Board Statement:** Not applicable. **Institutional Review Board Statement:** Not applicable.

**Appendix A** 

**Appendix A** 

**Data Availability Statement:** The data presented in this study are available on request from the corresponding author. **Informed Consent Statement:** Not applicable. **Data Availability Statement:** The data presented in this study are available on request from the **Informed Consent Statement:** Not applicable. **Data Availability Statement:** The data presented in this study are available on request from the

**Acknowledgments:** The authors extend gratitude to Bo Xu, Weiguo Li, Lin Wang and Hong Chang for their assistance in field data collection. The detailed comments from the editors and reviewers improved our study. corresponding author. **Acknowledgments:** The authors extend gratitude to Bo Xu, Weiguo Li, Lin Wang and Hong Chang for their assistance in field data collection. The detailed comments from the editors and reviewers corresponding author. **Acknowledgments:** The authors extend gratitude to Bo Xu, Weiguo Li, Lin Wang and Hong Chang for their assistance in field data collection. The detailed comments from the editors and reviewers

**Conflicts of Interest:** The authors declare no conflict of interest. improved our study. **Conflicts of Interest:** The authors declare no conflict of interest. improved our study. **Conflicts of Interest:** The authors declare no conflict of interest.

#### **Appendix A**

**Figure A1.** Predictive performance of the LSSVM regression models based on (**a**) all GLCM-based textures and (**b**) all multiscale textures for winter wheat AGB estimation. **Figure A1.** Predictive performance of the LSSVM regression models based on (**a**) all GLCM-based textures and (**b**) all multiscale textures for winter wheat AGB estimation. **Figure A1.** Predictive performance of the LSSVM regression models based on (**a**) all GLCM-based textures and (**b**) all multiscale textures for winter wheat AGB estimation.

**Figure A2.** Predictive performance of the PLSR model (**a**) and LSSVM regression model (**b**) using all spectra (400⁓1350 nm) for winter wheat AGB estimation. **Figure A2.** Predictive performance of the PLSR model (**a**) and LSSVM regression model (**b**) using all spectra (400⁓1350 nm) for winter wheat AGB estimation. **Figure A2.** Predictive performance of the PLSR model (**a**) and LSSVM regression model (**b**) using all spectra (400~1350 nm) for winter wheat AGB estimation.

#### **References References References**

1. Mistele, B.; Schmidhalter, U. Estimating the nitrogen nutrition index using spectral canopy reflectance measurements. *Eur. J. Agron.* **2008**, *29*, 184–190, doi:10.1016/j.eja.2008.05.007. 1. Mistele, B.; Schmidhalter, U. Estimating the nitrogen nutrition index using spectral canopy reflectance measurements. *Eur. J. Agron.* **2008**, *29*, 184–190, doi:10.1016/j.eja.2008.05.007. 1. Mistele, B.; Schmidhalter, U. Estimating the nitrogen nutrition index using spectral canopy reflectance measurements. *Eur. J. Agron.* **2008**, *29*, 184–190. [CrossRef]

