**3. Leaf car content assessment**

#### **3.1. Simulation results at the leaf level**

Based on PROSPECT-5 leaf simulations, correlation between CARI and LCar is presented in **Figure 1d**. Results showed that CARI had a significant linear relationship with LCar (R2 = 0.943, RMSE = 1.196 μg/cm<sup>2</sup> ), indicating that CARI index was accurate in estimating LCar with leaf-simulated data. Nevertheless, relationships between established spectral indices and LCar varied in **Figure 2**. Among these established spectral indices, the carotenoid indices, that is, CRI550, CRI700, CARred-edge and CARgreen, proposed by Gitelson et al. [11, 15] showed the highest correlation (R2 > 0.77, the RMSE < 2.40 μg/cm<sup>2</sup> ) with LCar. However, when LCar values were high, correlations between these indices and LCar presented with large dispersion, suggesting that these indices might be not sensitive to high LCar values (>15 μg/cm<sup>2</sup> ). Compared with CRI550 and CRI700, adding of a near infrared band (770 nm) in CARred-edge and CARgreen did not improve the estimation accuracy of LCar. Correlation between RARSc and LCar was general (R<sup>2</sup> = 0.603, RMSE = 3.160 μg/cm<sup>2</sup> ), and when LCar values were higher than 10 μg/cm2 , the correlation showed an obvious nonlinear trend. RBRI index was less correlated with LCar (R<sup>2</sup> = 0.165, RMSE = 4.584 μg/cm<sup>2</sup> ), and the scatter plot of RBRI versus LCar showed

**Figure 2.** Relationships between published spectral indices and leaf carotenoids content from leaf level data simulated

Monitoring Crop Carotenoids Concentration by Remote Sensing

http://dx.doi.org/10.5772/intechopen.78239

205

with PROSPECT-5.

**3. Leaf car content assessment**

**3.1. Simulation results at the leaf level**

the highest correlation (R2 > 0.77, the RMSE < 2.40 μg/cm<sup>2</sup>

LCar was general (R<sup>2</sup> = 0.603, RMSE = 3.160 μg/cm<sup>2</sup>

with LCar (R<sup>2</sup> = 0.165, RMSE = 4.584 μg/cm<sup>2</sup>

(R2 = 0.943, RMSE = 1.196 μg/cm<sup>2</sup>

10 μg/cm2

**Figure 1.** (a) R2

between CARI and LCar.

204 Progress in Carotenoid Research

Based on PROSPECT-5 leaf simulations, correlation between CARI and LCar is presented in **Figure 1d**. Results showed that CARI had a significant linear relationship with LCar

(b) Correlation between band 521 nm and LCar (c) Correlation between band 720 nm and LChl and (d) linear relationship

curves for LCar (LChl) versus leaf reflectance within the wavelength range from 400 to 800 nm

with leaf-simulated data. Nevertheless, relationships between established spectral indices and LCar varied in **Figure 2**. Among these established spectral indices, the carotenoid indices, that is, CRI550, CRI700, CARred-edge and CARgreen, proposed by Gitelson et al. [11, 15] showed

values were high, correlations between these indices and LCar presented with large dispersion, suggesting that these indices might be not sensitive to high LCar values (>15 μg/cm<sup>2</sup>

Compared with CRI550 and CRI700, adding of a near infrared band (770 nm) in CARred-edge and CARgreen did not improve the estimation accuracy of LCar. Correlation between RARSc and

, the correlation showed an obvious nonlinear trend. RBRI index was less correlated

), indicating that CARI index was accurate in estimating LCar

) with LCar. However, when LCar

), and when LCar values were higher than

), and the scatter plot of RBRI versus LCar showed

).

**Figure 2.** Relationships between published spectral indices and leaf carotenoids content from leaf level data simulated with PROSPECT-5.

large dispersity. PSSRc and PSNDc had low correlations with LCar. Compared with PSNDc, PSSRc showed a slightly better correlation with LCar (R<sup>2</sup> = 0.387, RMSE = 0.387 μg/cm<sup>2</sup> ). However, when LCar values exceeded 10 μg/cm<sup>2</sup> , PSSRc and PSNDc present obvious nonlinear correlations with LCar (**Figure 2g** and **h**). Correlation between PRI and LCar was poor (R2 = 0.120, RMSE = 4.705 μg/cm<sup>2</sup> ), and the scatter diagram showed obvious dispersion (**Figure 2i**). As for its modified version PRIm1, it showed almost no correlation with LCar, indicating that PRIm1 might not be suitable for the estimation of LCar. PSRI showed a low correlation with LCar (R<sup>2</sup> = 0.191, RMSE = 4.511 μg/cm<sup>2</sup> ), and the correlation was nonlinear. Different from these vegetation indices, SRcar showed a lower negative correlation with LCar (R2 = 0.142, RMSE = 4.645 μg/cm<sup>2</sup> ), and the scatter diagram also showed strong dispersity.

#### **3.2. Car assessment using ANGERS dataset**

First, the ANGERS dataset was used to analyze the capability of different spectral indices in estimating LCar. Performance of different spectral indices in LCar assessment is shown in **Table 3**. The estimation accuracy of CARred-edge and CARgreen in LCar assessment was slightly better than that of CRI550 and CRI700. However, compared with the simulated results, these indices showed rather poor performance in LCar retrieval with the ANGERS data. RARSc exhibited good performance in LCar retrieval with a R<sup>2</sup> value of 0.438 and a RMSE value of 3.792 μg/cm<sup>2</sup> . Although RBRI showed poor correlation with LCar in the leaf-simulated data, its estimation accuracy in LCar retrieval was the highest in the ANGERS data (R<sup>2</sup> = 0.727, RMSE = 2.640 μg/cm<sup>2</sup> ). Compared with PSNDc estimation results, estimation accuracy of PSSRc is relatively high, which is consistent with the foliar simulated results. PRI showed low accuracy in LCar estimation (R2 = 0.199, RMSE = 4.527 μg/cm<sup>2</sup> ), while PRIm1 also showed poor estimation results.

**Figure 3.** Scatterplots of measured LCar versus predicted LCar for spectral indices with ANGERS dataset. Dashed lines

CRI<sup>550</sup> 12 0.124 2.395 1.998 28.531 CRI<sup>700</sup> 13 0.046 2.533 2.121 30.171 CARgreen 7 0.411 1.941 1.637 23.122 CARred edge 9 0.344 2.050 1.739 24.417 **RARS***c* **2 0.674 1.443 1.130 17.192** RBRI 10 0.222 2.234 1.777 26.614 **PSND***c* **4 0.618 1.563 1.239 18.623** PSSR*c* 6 0.579 1.641 1.299 19.544 **PRI 1 0.710 1.369 1.092 16.305** PRIm1 11 0.125 2.373 1.814 28.268 PSRI 8 0.388 2.063 1.539 24.570 SRcar 5 0.614 1.571 1.144 18.713 **CARI 3 0.639 1.520 1.166 18.106**

**) MAE (μg/cm2**

Monitoring Crop Carotenoids Concentration by Remote Sensing

http://dx.doi.org/10.5772/intechopen.78239

207

**) RRMSE (%)**

indicate 1:1 lines.

**Index Rank R2 RMSE (μg/cm2**

**Table 4.** Cross-validation results for LCar estimation with wheat leaf level field data.


**Table 3.** Cross-validation results for LCar assessment using ANGERS data.

Monitoring Crop Carotenoids Concentration by Remote Sensing http://dx.doi.org/10.5772/intechopen.78239 207

large dispersity. PSSRc and PSNDc had low correlations with LCar. Compared with PSNDc, PSSRc showed a slightly better correlation with LCar (R<sup>2</sup> = 0.387, RMSE = 0.387 μg/cm<sup>2</sup>

linear correlations with LCar (**Figure 2g** and **h**). Correlation between PRI and LCar was

(**Figure 2i**). As for its modified version PRIm1, it showed almost no correlation with LCar, indicating that PRIm1 might not be suitable for the estimation of LCar. PSRI showed a low

Different from these vegetation indices, SRcar showed a lower negative correlation with LCar

First, the ANGERS dataset was used to analyze the capability of different spectral indices in estimating LCar. Performance of different spectral indices in LCar assessment is shown in **Table 3**. The estimation accuracy of CARred-edge and CARgreen in LCar assessment was slightly better than that of CRI550 and CRI700. However, compared with the simulated results, these indices showed rather poor performance in LCar retrieval with the ANGERS data. RARSc exhibited good per-

RBRI showed poor correlation with LCar in the leaf-simulated data, its estimation accuracy in LCar retrieval was the highest in the ANGERS data (R<sup>2</sup> = 0.727, RMSE = 2.640 μg/cm<sup>2</sup>

Compared with PSNDc estimation results, estimation accuracy of PSSRc is relatively high, which is consistent with the foliar simulated results. PRI showed low accuracy in LCar estima-

CRI<sup>550</sup> 10 0.139 4.693 3.363 54.179 CRI<sup>700</sup> 11 0.138 4.696 3.413 54.217 CARgreen 8 0.184 4.568 3.199 52.732 CARred-edge 7 0.190 4.550 3.232 52.524 **RARS***c* **3 0.438 3.792 2.757 43.781 RBRI 1 0.727 2.640 1.808 30.475** PSND*c* 9 0.167 4.617 3.472 53.303 **PSSR***c* **4 0.310 4.201 3.142 48.499** PRI 6 0.199 4.527 3.295 52.267 PRIm1 12 0.075 4.869 3.505 56.215 PSRI 13 0.002 5.057 3.796 58.377 SRcar 5 0.213 4.489 3.117 51.820 **CARI 2 0.545 3.413 2.345 39.400**

However, when LCar values exceeded 10 μg/cm<sup>2</sup>

correlation with LCar (R<sup>2</sup> = 0.191, RMSE = 4.511 μg/cm<sup>2</sup>

poor (R2 = 0.120, RMSE = 4.705 μg/cm<sup>2</sup>

206 Progress in Carotenoid Research

(R2 = 0.142, RMSE = 4.645 μg/cm<sup>2</sup>

formance in LCar retrieval with a R<sup>2</sup>

tion (R2 = 0.199, RMSE = 4.527 μg/cm<sup>2</sup>

**Index Rank R2 RMSE (μg/cm2**

**Table 3.** Cross-validation results for LCar assessment using ANGERS data.

**3.2. Car assessment using ANGERS dataset**

).

. Although

).

, PSSRc and PSNDc present obvious non-

), and the correlation was nonlinear.

), and the scatter diagram showed obvious dispersion

), and the scatter diagram also showed strong dispersity.

value of 0.438 and a RMSE value of 3.792 μg/cm<sup>2</sup>

**) MAE (μg/cm2**

), while PRIm1 also showed poor estimation results.

**) RRMSE (%)**

**Figure 3.** Scatterplots of measured LCar versus predicted LCar for spectral indices with ANGERS dataset. Dashed lines indicate 1:1 lines.


**Table 4.** Cross-validation results for LCar estimation with wheat leaf level field data.

Among all the indices, PSRI had the lowest estimation accuracy, possibly due to its insensitive to LCar. The estimation accuracy of SRcar generally (R<sup>2</sup> = 0.213, RMSE = 4.489 μg/cm<sup>2</sup> ) ranks fifth in all estimation results. Compared with these existing spectral indices, the estimation accuracy of CARI was accurate (R<sup>2</sup> = 0.545, RMSE = 0.545 μg/cm<sup>2</sup> ), second to RBRI, showing that CARI data can be used to accurately estimate LCar in the ANGERS data.

Based on the estimation results of these spectral indices in LCar retrieval with the ANGERS dataset, the scatter diagrams of the best four ranking spectral indices were presented in **Figure 3**. The results showed that compared with other indices, the fitting line of the scatterplot of RBRI is closer to the 1:1 straight line (the slope of the fitting line is 0.730).In addition, RBRI index was more sensitive to higher leaf carotenoid content (>15 μg/cm<sup>2</sup> ). The CARI index also showed good estimation results, except for the samples that had LCar values greater than 15 μg/cm2 , and the estimated values of most sample points were evenly distributed around the 1:1 straight line with the measured values. Compared with RBRI, CARI was more sensitive to lower LCar values (<3 μg/cm<sup>2</sup> ), but it showed a slight "saturation effect" on high LCar values (>15 μg/cm2 ). RARS and PSSRc indices also showed satisfactory estimation results. Similar to CARI, these indices were not sensitive to higher LCar values (>15 μg/cm<sup>2</sup> ).
