*3.3. E*ff*ects of Di*ff*erent Segmentation Methods on Maize Growth Dynamics*

#### 3.3.1. Spectral Reflectance Response

Figure 6 lists the maize canopy spectral reflectance trends extracted from the original spectral images and the three segmentation methods obtained during the three growth periods. In general, the spectral reflectance of the maize canopy decreases in the blue and red bands and increases in the green, red edge, and near-infrared bands after the multispectral image is segmented. This phenomenon presents the typical spectral feature of green vegetation with light absorption by chlorophyll and carotene pigment in blue and red bands, and relative higher light reflectance in green and NIR ranges. The effects of the three segmentation methods in different growth periods in Figure 6 vary. As shown

in Figure 6a, the maize canopy coverage at the seedling stage is low, the soil background is highly visible, and the segmentation is easy; thus, the results obtained by the three segmentation methods are similar. The maize grows to the jointing and ear stages, and the maize canopy structure is complex, which is not conducive to background culling. Figure 6b,c show that the reflectance of ExG index segmented image is lower than image segmented by the wavelet or threshold methods. However, the performances of wavelet and threshold segmentation are very close in all three stages.

The segmentation results are difficult to evaluate by ideal images, which are without influence factors in the field. We therefore compared the segmented image with the original image to explore the differences between three methods. Table 4 shows results of RMSE operated by three segmentation methods. At the seeding stage, the RMSE values at NIR bands are higher than other band, and decrease following ExG, threshold and wavelet segmentation in blue, green, red, and red edge bands; the value of wavelet segmentation method is lowest in each band. At jointing and ear stage, RMSE values of threshold and wavelet segmentation are close and higher than ExG segmentation results, especially for green and NIR bands. Meanwhile, RMSE value of the wavelet method is higher than that of the threshold method in each band. Higher values show that changes and quality improvement between segmented image and original image in each band are more significant. Thus, the wavelet segmentation method shows a better performance for field difference enhancement and image quality improvement.


**Table 4.** Comparison of root mean square error (RMSE) between different segmentation and original image.

**Figure 6.** Calculation of maize canopy spectral reflectance by different segmentation methods. **Figure 6.** Calculation of maize canopy spectral reflectance by different segmentation methods.

#### 3.3.2. Texture Response 3.3.2. Texture Response

achieved stable results in the three growth stages.

The maize canopy image in the multispectral image was extracted by background culling to generate a maize canopy vector file. We wanted to further discuss texture responses in processed images. The texture change trend of the maize canopy in different growth periods was extracted by various segmentation methods. Figure 7 demonstrates that the standard deviation, smoothness, and entropy of the near-infrared image of the maize canopy gradually increase with the growth of the maize when the maize canopy is not segmented. This phenomenon indicates that the canopy The maize canopy image in the multispectral image was extracted by background culling to generate a maize canopy vector file. We wanted to further discuss texture responses in processed images. The texture change trend of the maize canopy in different growth periods was extracted by various segmentation methods. Figure 7 demonstrates that the standard deviation, smoothness, and entropy of the near-infrared image of the maize canopy gradually increase with the growth of the

7 shows that compared with ExG index segmentation method, threshold segmentation method and wavelet segmentation method can better reduce the standard deviation, smoothness, and entropy of the image. Firstly, the results of threshold segmentation method were higher than others. Even the standard deviation and entropy valued of the threshold segmented image are close to the image segmented by wavelet and ExG, respectively, its performances are not the best. Secondly, the standard deviation of ExG processed image is lowest at the seedling stage due to the soil background removal; the results of standard deviation and smoothing are close between ExG and wavelet segmentation methods at the jointing and the ear stage. When compared to the wavelet method, although the ExG method has lower values of standard deviation and smoothness, the entropy value of ExG segmented image is close to threshold segmentation and higher than wavelet segmentation at the jointing and the ear stage. It means that the randomness performances processed by threshold and ExG segmentation are higher than wavelet segmentation because of the limited elimination of canopy structure influences during these stages. Thus, the wavelet segmentation method has

structure of the plant canopy became more and more complicated with the growth of the maize.

maize when the maize canopy is not segmented. This phenomenon indicates that the canopy structure of the plant canopy became more and more complicated with the growth of the maize.

Three segmentation methods were used to eliminate influences form the soil background, all of which reduce the standard deviation, smoothness, and entropy of the image to a certain extent. Figure 7 shows that compared with ExG index segmentation method, threshold segmentation method and wavelet segmentation method can better reduce the standard deviation, smoothness, and entropy of the image. Firstly, the results of threshold segmentation method were higher than others. Even the standard deviation and entropy valued of the threshold segmented image are close to the image segmented by wavelet and ExG, respectively, its performances are not the best. Secondly, the standard deviation of ExG processed image is lowest at the seedling stage due to the soil background removal; the results of standard deviation and smoothing are close between ExG and wavelet segmentation methods at the jointing and the ear stage. When compared to the wavelet method, although the ExG method has lower values of standard deviation and smoothness, the entropy value of ExG segmented image is close to threshold segmentation and higher than wavelet segmentation at the jointing and the ear stage. It means that the randomness performances processed by threshold and ExG segmentation are higher than wavelet segmentation because of the limited elimination of canopy structure influences during these stages. Thus, the wavelet segmentation method has achieved stable results in the three growth stages. *Remote Sens.* **2020**, *12*, x FOR PEER REVIEW 13 of 20

**Figure 7.** Calculation of the maize canopy texture features using different segmentation methods. **Figure 7.** Calculation of the maize canopy texture features using different segmentation methods.

#### 3.3.3. Coverage Response 3.3.3. Coverage Response

three segmentation methods.

discussion part.

Figure 8 shows the change trend of maize plant coverage calculated by extracting the canopy area by using three segmentation methods during the above-mentioned three growth stages. The maize canopy coverage was calculated by ratio of segmented canopy pixels and full image pixels and presented between 0 and 1. In general, the values of canopy coverage acquired by all of segmentation methods increased with the growth stages. Some differences in the maize canopy coverage calculated by different segmentation methods can be observed in the same growth period. Firstly, canopy coverage values calculated based on three segmented results at the seedling stage were relatively small and lower than 0.45. This phenomenon is attributed to the simple structure of the maize plant at the seedling stage. The visible part of the soil is large, and the division is simple when the leaf volume is low. Secondly, those at the jointing and ear stages were different, especially ExG segmentation results that were significantly higher than others above 0.75; the coverage values of threshold and wavelet segmented images were similar and higher than 0.45; the result of the threshold segmentation was a little bit higher than wavelet segmentation method at the earing stage. The significantly differences between three segmentation methods at the jointing and ear stages also show that the segmentation was more complicated than the seedling stage due to a large number of leaves crossing each other and a complex canopy structure of maize plant. As mentioned above, in the three segmentation methods used in this study, the maize canopy Figure 8 shows the change trend of maize plant coverage calculated by extracting the canopy area by using three segmentation methods during the above-mentioned three growth stages. The maize canopy coverage was calculated by ratio of segmented canopy pixels and full image pixels and presented between 0 and 1. In general, the values of canopy coverage acquired by all of segmentation methods increased with the growth stages. Some differences in the maize canopy coverage calculated by different segmentation methods can be observed in the same growth period. Firstly, canopy coverage values calculated based on three segmented results at the seedling stage were relatively small and lower than 0.45. This phenomenon is attributed to the simple structure of the maize plant at the seedling stage. The visible part of the soil is large, and the division is simple when the leaf volume is low. Secondly, those at the jointing and ear stages were different, especially ExG segmentation results that were significantly higher than others above 0.75; the coverage values of threshold and wavelet segmented images were similar and higher than 0.45; the result of the threshold segmentation was a little bit higher than wavelet segmentation method at the earing stage. The significantly differences between three segmentation methods at the jointing and ear stages also show that the segmentation was more complicated than the seedling stage due to a large number of leaves crossing each other and a complex canopy structure of maize plant.

and the presence of many background effects, such as soil and shadow. The results are difficult to prove by actual coverage values because of the limitation of ideal image acquisition, so that we resort to the diagnosis modeling of chlorophyll content to further estimate the capability and validity of

However, the phenomena presented in Figure 8 could be discussed based on the combination of the spectral response in the field and the principal of each segmentation method. The detail is in

coverage calculated by the ExG vegetation index method was high. Meanwhile, the maize canopy coverage calculated by the threshold and wavelet segmentation methods was low. This study takes

As mentioned above, in the three segmentation methods used in this study, the maize canopy coverage calculated by the ExG vegetation index method was high. Meanwhile, the maize canopy coverage calculated by the threshold and wavelet segmentation methods was low. This study takes the ear period as an example. The maize coverage calculated by the ExG vegetation index method, threshold segmentation method, and wavelet segmentation method was 0.764, 0.562, and 0.520, respectively. This result is attributed to the complicated canopy structure at the ear stage of maize and the presence of many background effects, such as soil and shadow. The results are difficult to prove by actual coverage values because of the limitation of ideal image acquisition, so that we resort to the diagnosis modeling of chlorophyll content to further estimate the capability and validity of three segmentation methods.

However, the phenomena presented in Figure 8 could be discussed based on the combination of the spectral response in the field and the principal of each segmentation method. The detail is in *Remote Sens.*  discussion part. **2020**, *12*, x FOR PEER REVIEW 14 of 20

#### *3.4. Modeling and Analysis of Maize Canopy Chlorophyll Content 3.4. Modeling and Analysis of Maize Canopy Chlorophyll Content*

On the basis of the above analysis, the maize canopy multispectral image is segmented, and the background interference in the image is eliminated. The multispectral band reflectance, multiple vegetation indices, and maize canopy chlorophyll content were established to construct a PLS model. Correlation analysis was performed using the maize canopy chlorophyll content calculated by the model and the real value measured in the experiment. The maize canopy chlorophyll based on different segmentation methods was established by comparing the R2 and RMSE of the two groups of variables. The test model was analyzed, and the results are shown in Table 5. Figure 9 shows scatter plots of the predicted and true values obtained under different segmentation methods to intuitively display their relationship. On the basis of the above analysis, the maize canopy multispectral image is segmented, and the background interference in the image is eliminated. The multispectral band reflectance, multiple vegetation indices, and maize canopy chlorophyll content were established to construct a PLS model. Correlation analysis was performed using the maize canopy chlorophyll content calculated by the model and the real value measured in the experiment. The maize canopy chlorophyll based on different segmentation methods was established by comparing the R<sup>2</sup> and RMSE of the two groups of variables. The test model was analyzed, and the results are shown in Table 5. Figure 9 shows scatter plots of the predicted and true values obtained under different segmentation methods to intuitively display their relationship.

Table 5 shows the model accuracy of the reflectivity modeling obtained using the original reflectivity modeling and the three segmentation methods. The model built using the original reflectivity has low accuracy and large errors. Compared with the original model, the model constructed based on the reflectance obtained by the three segmentation methods has improved performance to a certain extent in terms of accuracy and error. This indicates that removing the soil background from the original image can improve the diagnostic accuracy of chlorophyll content. Compared with the ExG index segmentation method and the threshold segmentation method, the model based on the wavelet segmentation method shows the best performance and the lowest error (Rc2 = 0.6638, RMSEF = 3.6211, MAE = 2.89; Rv2 = 0.6923, RMSEP = 3.9067, and MAE = 3.19). Table 5 shows the model accuracy of the reflectivity modeling obtained using the original reflectivity modeling and the three segmentation methods. The model built using the original reflectivity has low accuracy and large errors. Compared with the original model, the model constructed based on the reflectance obtained by the three segmentation methods has improved performance to a certain extent in terms of accuracy and error. This indicates that removing the soil background from the original image can improve the diagnostic accuracy of chlorophyll content. Compared with the ExG index segmentation method and the threshold segmentation method, the model based on the wavelet segmentation method shows the best performance and the lowest error (Rc<sup>2</sup> = 0.6638, RMSEF = 3.6211, MAE = 2.89; Rv<sup>2</sup> = 0.6923, RMSEP = 3.9067, and MAE = 3.19).

With background 0.5431 4.2184 3.24 0.5894 4.6947 3.63 Threshold segmentation 0.6330 3.7270 2.93 0.6302 4.5248 3.49 ExG index segmentation 0.6584 3.8512 3.07 0.5601 4.6992 2.96 Wavelet segmentation 0.6638 3.6211 2.89 0.6923 3.9067 3.19

**Table 5.** Diagnostic results of maize canopy chlorophyll with different segmentation methods.



**Figure 9.** Maize canopy chlorophyll test results. **Figure 9.** Maize canopy chlorophyll test results.

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

the soil background.

Spectral analysis and imaging spectroscopy, a fast and nondestructive diagnosis method, can comprehensively reflect the changes of physiological and biochemical indexes and canopy structure of crops by analyzing the spectral reflectance characteristics of crops; accordingly, a rapid nondestructive diagnosis of crop nutrition is achieved [6,7]. The reflectance spectral data can be used to diagnose the chlorophyll content of maize in the field. The image acquired by a UAV-based spectral platform contains rich information, including not only spectral information of ground objects but also texture and coverage data. Spectral analysis and imaging spectroscopy, a fast and nondestructive diagnosis method, can comprehensively reflect the changes of physiological and biochemical indexes and canopy structure of crops by analyzing the spectral reflectance characteristics of crops; accordingly, a rapid nondestructive diagnosis of crop nutrition is achieved [6,7]. The reflectance spectral data can be used to diagnose the chlorophyll content of maize in the field. The image acquired by a UAV-based spectral platform contains rich information, including not only spectral information of ground objects but also texture and coverage data.

#### *4.1. Elimination of the Impact of Background and Noise 4.1. Elimination of the Impact of Background and Noise*

This study showed that the spectral reflectance of maize canopy greatly changed after the soil background was removed, which was consistent with the results of Wu et al. [22]. Compared with the original image with the soil background, the spectral reflectance values reduce in blue and red bands, and increase in green, red edge, and NIR bands. It is similar to that of the maize canopy in Figure 4. The gray value of the maize canopy is lower than that of the soil in the blue and red bands This study showed that the spectral reflectance of maize canopy greatly changed after the soil background was removed, which was consistent with the results of Wu et al. [22]. Compared with the original image with the soil background, the spectral reflectance values reduce in blue and red bands, and increase in green, red edge, and NIR bands. It is similar to that of the maize canopy in Figure 4. The gray value of the maize canopy is lower than that of the soil in the blue and red bands

Figure 7 shows that the texture features (standard deviation, smoothness, and entropy) of maize canopy multispectral images become more complex with the extension of growth period. The main reason is that with the growth of maize, the area and number of leaves increase and cross each other, which increases the complexity of canopy structure. Three segmentation methods are used to segment the soil background in the multispectral image, and only the corn part of the image is preserved, which reduces the type of elements in the image and reduces the complexity of the image. and is higher than that of the soil in the green, red, and near-infrared bands. It shows that the canopy segmentation could help to enhance the spectral characteristic of crop and reduce the influences from the soil background.

Figure 7 shows that the texture features (standard deviation, smoothness, and entropy) of maize canopy multispectral images become more complex with the extension of growth period. The main reason is that with the growth of maize, the area and number of leaves increase and cross each other, which increases the complexity of canopy structure. Three segmentation methods are used to segment the soil background in the multispectral image, and only the corn part of the image is preserved, which reduces the type of elements in the image and reduces the complexity of the image. Therefore, the standard deviation, smoothness, and entropy in Figure 7 are reduced to some extent by three segmentation methods. Compared with threshold segmentation method and wavelet segmentation method, ExG exponential segmentation method has lower performance. The possible reason is that the illumination change during UAV image acquisition interferes with the calculation of vegetation index, which is easily saturated under high biomass conditions. Previous studies have shown that wavelet transform can be used to process UAV images and decompose different proportions of image signals to eliminate noise [38]. The wavelet segmentation method combines the advantages of wavelet transform and threshold segmentation, and can achieve good results in noise elimination and object classification and classification [39]. This is also confirmed by the results of this study. The wavelet segmentation method is used to remove the soil background, reduce the complexity of crop canopy, and achieve stable performance in three growth periods. Figure 8 shows the results of the three segmentation methods to calculate the coverage of maize at three growth stages. Although the ExG segmentation method is widely used in the field crop detection, it is operated based on different spectral bands. According to the spectral characteristics of crop and soil background, the ExG index shown in Equation (9) is established by grey values of blue, green, and red bands to amplify the spectral difference between crop and soil background. As shown in Figure 4, this way pays attention to the light absorption by chlorophyll and carotene pigment in visible bands related to blue and red, it reduces the soil background which has close grey values among blue, green, and red band image. While threshold and wavelet segmentation methods are applied in NIR image in which the grey values are significant different between corn and soil to show the biological vitality. Compared with the NIR image, grey values are easily affected by environment in visible bands. For example, a strong light radiation might reduce grey contrast and lead miss pixels of leaves, the canopy coverage might be lower than those of threshold and wavelet segmentation methods at the seedling stage. At the jointing and ear stage, the coverage values calculated by ExG show significant changes, because the soil background is misclassified as the crop due to more complex structure of plants and the expanded shading elements from the crossing leaves. The coverage value at ear stage is lower than those at jointing stage because of closed crop lines by luxuriant leaves and reduced misclassification of soil back ground.

Regards to the segmented results of the threshold and wavelet method, the texture analysis helps to evaluate the segmented results and explain how it conducts to improve the quality from image contrast, brightness, and randomness. It is discussed based on texture analysis results shown in Figure 7a,b; standard deviation and smoothness values of wavelet segmented images are higher than that of threshold segmented image at seedling stage, lower at ear stage, and similar at jointing stage, respectively. The higher value of standard deviation indicates the higher image contrast, and it means the edge pixels of leaves are more clearly to classify. The higher smoothness value indicates the lower influence from noise or background, and it shows better performances of image enhancement. As a result, the coverage result of the wavelet segmentation method is higher than that of the threshold segmented image at seedling stage, lower at ear stage, and similar at jointing stage, respectively.

In addition, the wavelet analysis is a method usually used in signal denoising and enhancement. Figure 7c shows that the randomness indicated by entropy is improved by wavelet method during jointing and ear stages. It is contributed by the image denoising to reduce the effects of canopy structure

and leaves crossing. The wavelet method also helps to enhance useful edge or detail information in the NIR image so that the randomness locates between ExG and the threshold method at seedling stage. From this view, the wavelet segmentation method not only eliminates the soil background in the image, but also segments the edge and noise points in the image [38], so the calculated coverage by wavelet method is more reliable than the ExG index and threshold segmentation method.

#### *4.2. Diagnosis of Chlorophyll Content in Maize Based on Di*ff*erent Segmentation Methods*

The vegetation index has been widely used to diagnose crop nutritional parameters, and its performance varies with wavelength configuration [23,52]. Duan et al. [53] found saturation in estimating LAI (Leaf Area Index) by using NDVI alone. Cheng et al. [16] found that the four NIR-based vegetable indices outperformed the two-color indices in AGB estimation across all growth stages. Therefore, a single band or vegetation index was not easy to show good model accuracy in all growth periods. In this study, 20 commonly used vegetation indices were selected to construct a chlorophyll content diagnostic model for the entire growth period of maize.

The spectral reflectance of maize canopy was obtained on the basis of the three segmentation methods to remove soil background, and the vegetation index was calculated. A chlorophyll content diagnostic model of maize canopy based on different segmentation methods was constructed using the PLS method. The results showed that the accuracy of the diagnostic model constructed using the original canopy spectral data was low (Rc<sup>2</sup> = 0.5431, RMSEF = 4.2184, MAE = 3.24; Rv<sup>2</sup> = 0.5894, RMSEP = 4.6947, and MAE = 3.36), and the accuracy of the diagnosis of chlorophyll content in maize canopy could be improved to varying degrees by using three segmentation methods to remove background noise in maize multispectral images. However, among the three segmentation methods selected in this study, the accuracy of the training set based on the ExG index segmentation method was 0.6584, but the accuracy of the verification set was only 0.5601, which indicates that the model is not robust enough. This result is consistent with the previous discussion, and the color index is easily affected by ambient light. In the future, spectral radiometer can be used to correct the reflectivity of UAV images, so as to further optimize the performance of the model. The diagnostic model of maize canopy chlorophyll content based on wavelet segmentation had the highest quasi determination (Rc<sup>2</sup> = 0.6638, RMSEF = 3.6211, MAE = 2.89; Rv<sup>2</sup> = 0.6923, RMSEP = 3.9067, and MAE = 3.19). This also confirmed that the wavelet segmentation method can effectively remove the soil background and noise points in the UAV multispectral image and restore the spectral reflection characteristics of crops in the canopy image. The good performance of wavelet segmentation also shows that image enhancement has great potential in UAV image noise elimination.

#### **5. Conclusions**

In the present study, we studied the background noise in the multispectral images of maize canopy acquired by UAV and discussed the performance of three noise rejection methods. The results show that the soil background removal from multispectral images can reduce the complexity of image texture and improve the spectral reflectance characteristics of crop canopy. The wavelet segmentation method achieved better noise rejection effect in the three growth stages of maize compared with the threshold and ExG index segmentation methods. We extracted 20 common vegetation indices on the basis of the original spectral images and the spectral images after background removal and constructed a diagnostic model of maize canopy chlorophyll content using the partial least squares regression method. The results showed that the removal of the soil background was helpful in improving the diagnostic accuracy of the maize canopy chlorophyll content in the field. The diagnostic model of the chlorophyll content in maize canopy based on wavelet segmentation achieved the optimal results (Rc<sup>2</sup> = 0.6638, RMSEF = 3.6211; Rv<sup>2</sup> = 0.6923, and RMSEP = 3.9067). With the popularity of the UAV remote sensing platform in recent years, the results of this study can provide a valuable reference for noise removal during UAV image crop nutrition monitoring. However, this study was based on

small-area plot experiments, and other noises besides soil background still exist in field experiments. Such issue was not discussed in depth in this study. Future work will be conducted in this direction.

**Author Contributions:** Conceptualization, L.Q., D.G., J.Z., M.L., H.S. and J.M.; Data curation, L.Q., D.G., J.Z., M.L., H.S. and J.M.; Formal analysis, L.Q., D.G., J.Z., M.L., H.S. and J.M.; Funding acquisition, L.Q., J.Z., M.L., H.S. and J.M.; Investigation, L.Q., D.G., J.Z., M.L., H.S. and J.M.; Methodology, L.Q., D.G., J.Z., M.L., H.S. and J.M.; Project administration, M.L., H.S. and J.M.; Resources, L.Q., D.G., J.Z., M.L., H.S. and J.M.; Software, L.Q., D.G., J.Z., M.L. and H.S.; Supervision, L.Q., D.G., J.Z., M.L., H.S. and J.M.; Validation, L.Q., D.G. and M.L.; Visualization, L.Q., J.Z. and H.S.; Writing—original draft, L.Q.; Writing—review & editing, L.Q., M.L., H.S. and J.M. All authors have read and agreed to the published version of the manuscript.

**Funding:** The project was supported by the National Key Research and Development Program (2018YFD0300505-1), the National Natural Science Fund (Grant No. 31971785 and 31501219), the Fundamental Research Funds for the Central Universities (Grant No. 2020TC036), and the Graduate Training Project of China Agricultural University (JG2019004 and YW2020007).

**Acknowledgments:** We would like to thank Zizheng Xing, Zhiyong Zhang, Xuying Ma, Di Song, Ruomei Zhao, and Song Li for their help with field data collection.

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