3.1. Evaluation of Data Acquisition Performance of the Portable Three-Band CGMD Instrument
To evaluate the performance of the CGMD instrument in obtaining vegetation indices, regression analysis was performed on the corresponding values of NDVI, RVI, DVI, TVI obtained by the CGMD instrument in each growth period. To facilitate the comparison and analysis of the fitting accuracy of the combined vegetation indices of different wavebands, the vegetation indices obtained by FieldSpec HandHeld2 and CGMD were normalized. As shown in Figure 5
, the fitting results of the vegetation indices obtained by CGMD and the corresponding values obtained by the commercial instrument FieldSpec HandHeld2 have a linear relationship, but some differences can be seen in the fitting accuracy of different vegetation.
The correlation of fitting of each vegetation index with different band combinations was compared. The fitting results of NDVI (R730, R815) were better than those of NDVI (R660, R815) and NDVI (R660, R730). The R2 values were 0.83, 0.67, and 0.56, respectively, and the RRMSE values were 0.16, 0.10, and 0.13, respectively; the combination of 660 and 730 nm and 660 and 815 nm showed severe dispersion at low values. There were similar patterns for different band combinations for RVI and DVI. For the 660 nm/730 nm, 660 nm/815 nm, and 730 nm/815 nm combinations, the R2 values of RVI were 0.53, 0.66, and 0.82, respectively, and the RRMSE values were 0.33, 0.35, and 0.07, respectively; the R2 values of DVI were 0.21, 0.60, and 0.79, respectively, and the RRMSE values were 0.21, 0.60, and 0.79, respectively. The accuracies of vegetation indices of different bands obtained by CGMD were in a descending order of 730 nm/815 nm, 660 nm/815 nm, and 660 nm/730 nm. This result was caused by the accuracy difference of the sensor at each band. The three-band vegetation indices TVI-1 and TVI-2 obtained by CGMD also showed a relatively high accuracy, with R2 values of 0.78 and 0.69, respectively, and RRMSE values of 0.10 and 0.27, respectively. The correlation of fitting of CGMD with FieldSpec HandHeld2 was relatively high, the R2 values were above 0.5 (except for DVI (R660, R730)), and CGMD could be used to obtain wheat canopy reflectance.
3.3. Spectral Monitoring Model of Wheat Growth
From the fitting of the vegetation index with the growth index, under the condition that the wheat canopy population was large, the vegetation index was saturated, which led to poor prediction accuracy of each vegetation index with high biomass. As shown in Figure 6
, Figure 8
and Figure 10
, as the value of growth indices increased, the distribution of scattered points became more discrete, and this problem was difficult to solve with a single spectral monitoring model. According to the pattern of wheat growth, various growth indices in the vegetative growth stage tended to increase rapidly, and after the shift from vegetative growth to reproductive growth, the trend of increase was flat, and some indices even decreased. Based on the differences of wheat growth indices in different phenological periods, the samples could be classified, which could reduce the gap between the values of scattered points with high biomass in each model. Therefore, this paper proposes a method for constructing a wheat-growth-index spectral monitoring model based on the growth stages. The data were divided into periods I and II according to the characteristics of wheat growth, and a growth spectral monitoring model was independently constructed to solve the problem of impact of vegetation index saturation on the prediction accuracy of the model.
The dynamic changes of wheat growth indices from the jointing stage to the flowering stage are shown in Figure 14
. The value of LAI before the heading stage increased over the growth period; the increase rate from the jointing stage to the booting stage was relatively small, the average value increased from 1.75 to 2.06, and the value ranges were 0.55–2.87 and 0.69–3.88, respectively. The value from the booting stage to the heading stage varied greatly; the average value increased to 2.80, and the value at the heading stage ranged from 0.80 to 5.48. After the heading stage, the wheat LAI value changed slightly, the average value at the flowering stage was 2.98, and the value range was 1.02–5.54. Therefore, the period from the jointing stage to the booting stage was defined as period I, and the period from the heading stage to the flowering stage was defined as period II.
The LDW value of wheat increased greatly from the jointing stage to the booting stage; the average value increased from 112.13 g/m2 to 191.48 g/m2, and the value range increased from 36.56–199.52 g/m2 to 53.35–371.49 g/m2. The value from the booting stage to the flowering stage was relatively stable, the value from the booting stage to the heading stage decreased, and the average value decreased from 191.48 g/m2 to 144.05 g/m2. The value range at the booting stage was 27.83–302.18 g/m2. The value from the booting stage to the flowering stage remained the same. The average value at the flowering stage was 143.03 g/m2, with a range of 46.64–330.31 g/m2. Therefore, the jointing stage alone was defined as period I, and the booting stage to the flowering stage was defined as period II.
The differences of LNC in the wheat phenological stages were small, and the average values from the jointing stage to the flowering stage were 2.93%, 2.77%, 2.82%, and 2.84%, respectively. The range of the jointing stage was close to that of the booting stage, and the value range was mostly distributed at 1.7%–3.5%, while the value ranges of the heading stage and the flowering stage were more consistent, and the value range was mostly distributed at 1.8%–4.0%. Therefore, the jointing stage and the booting stage were defined as period I, and the heading stage and flowering stage were defined as period II.
LNA is a growth index obtained by multiplying LNC and LDW. Its trend of change includes the characteristics of LDW and LNC. There was a large change in the value from the jointing stage to the booting stage, the average value increased from 3.45 g/m2 to 5.70 g/m2, and the range increased from 0.64–7.13 g/m2 to 0.94–13.27 g/m2. The values from the booting stage to the flowering stage tended to be stable. The average values were 5.70 g/m2, 4.47 g/m2, and 4.51 g/m2, respectively, and the values ranged from 0.94–13.27 g/m2, 0.55–10.00 g/m2, and 1.03–11.97 g/m2, respectively. Therefore, the jointing stage alone was defined as period I, and the period from the booting stage to the flowering stage was defined as period II.
The vegetation indices with the highest prediction accuracy corresponding to different wheat growth indices of periods I and II were screened out to construct the crop growth monitoring model (Figure 15
). The number of samples used for constructing models and the verification was 63 and 28, respectively. The vegetation index corresponding to LAI with the highest prediction accuracy was TVI-2, the vegetation index corresponding to LDW was TVI-1, the vegetation index corresponding to LNC was NDVI (R730
), and the vegetation index corresponding to LNA was NDVI (R730
). The monitoring model and verification results are shown in Figure 15
. The fitting accuracies of LAI, LDW, and LNA in period II increased, R2
values increased to 0.94, 0.85, and 0.87, respectively, and the RRMSE values were 0.13, 0.08, and 0.32, respectively, showing that the saturation problem caused by large groups in the late growth stage was eliminated to some extent. However, the fitting result of period I was not considerably improved compared with that of the original monitoring model, and the quality of fitting results of the growth indices, except for LAI and LNC, decreased to a certain extent. The reason may be that the data range and coefficient of variation of the early growth stage are small, and under the same conditions, the required accuracy of the monitoring instrument is higher. Additionally, the reduction in the number of training samples for modeling in different periods also has a certain impact on the accuracy of the model. Therefore, in this study, we still chose the monitoring model of the whole growth period when predicting the values of the growth indices LDW and LNA in period I. The trend lines, R2
values, and RRMSE values of LNC in period I and period II were consistent, indicating that the change in the numerical characteristics of LNC in different phenological periods is small, and segmentation modeling cannot improve the accuracy of LNC fitting.
The verification results showed that the prediction accuracy of the LAI monitoring model increased the most, R2 increased from 0.64 to 0.79, and RRMSE decreased from 0.29 to 0.22. The prediction accuracy improved to a certain extent when the population was large, and the trend line was closer to the 1:1 line. The prediction accuracies of LDW and LNA improved slightly, the R2 values increased to 0.85 and 0.85, respectively, and the RRMSE values were 0.23 and 0.28, respectively. The prediction result of LNC was the same as that of the original monitoring model, and the segmentation modeling could not improve its monitoring accuracy. The above finding shows that the wheat growth monitoring model constructed based on the growth stages in this study greatly improves the monitoring accuracy of LAI, but the prediction accuracy of LDW, LNC, and LNA were not able to be substantially improved.