**3. Results**

## *3.1. In-Situ Data*

Table 2 presents descriptive statistics for the measured concentrations of Chl-a and TSM in the water samples obtained from the three samplings on Poyang Lake.

**Table 2.** Descriptive statistics for the measured concentrations of chlorophyll-a (Chl-a) and total suspended matter (TSM) in Poyang Lake.


Note: SD is the standard deviation; C.V. is the coefficient of variation.

The water reflectance spectra were obtained by calculating the radiance that was collected at the Poyang Lake sampling sites. After removing two abnormal spectral data (with extremely low reflectance) in August 2015 sampling, a total of 33 valid spectra were obtained for the August 2015 sampling, 17 valid spectra for the October 2015 sampling, and 14 valid spectra for the January 2016 sampling (Figure 4). The reflectance spectra show a reflectance peak between 550 and 600 nm, which is related to the weak absorption of chlorophyll and carotene. The absorption of cyanophycin leads to an absorption valley between 600 and 650 nm; however, this valley was not observed in the reflectance spectra in this study due to the low concentration of chlorophyll in Poyang Lake. However, we observed an obvious absorption between 660 and 680 nm, which is caused by chlorophyll in the red-light band. The absorption becomes more obvious as the concentration of chlorophyll increases. We also observed a reflection peak near 700 nm. This reflection peak is an important feature of algae-containing water, and its position and amplitude indicate the concentration of Chl-a, with the peak moving towards longer wavelengths as the concentration of Chl-a increases. The reflectance of the spectra is low after about 730 nm, which is due to the strong absorption of Chl-a in the near-infrared band. A small reflection peak appears near 820 nm, which may be due to the scattering of suspended matter. For some sampling sites, the peaks and valleys in the reflectance spectra are not strongly pronounced, which is mainly due to the low concentration of Chl-a and the high concentration of TSM. The measured reflectance spectra vary for different seasons due to the change in the lake's water area throughout the year and the consequent change in the concentrations of Chl-a and TSM.

**Figure 4.** Measured water reflectance spectra for (**a**) August 2015, (**b**) October 2015, and (**c**) January 2016.

Of all the sampling sites that were used in this study, two sets of sampling sites were selected to determine the characteristic reflectance bands of Chl-a and TSM. Figure 5a shows reflectance spectra from three sampling sites with approximately the same Chl-a concentration (~4 mg/m3) and different TSM concentrations (18.6 mg/L, 55.8 mg/L, and 98.4 mg/L). Meanwhile, Figure 5b shows reflectance spectra from three sampling sites with similar TSM concentrations (~6 mg/L) and different Chl-a concentrations (3.91 mg/m3, 6.36 mg/m3, and 9.70 mg/m3).

**Figure 5.** (**a**) The measured remote-sensing-reflectance spectra of the surface of Poyang Lake for sites with similar concentrations of Chl-a and different concentrations of TSM. (**b**) The measured remote-sensing-reflectance spectra of the surface of Poyang Lake for sites with similar concentrations of TSM and different concentrations of Chl-a.

From the spectra shown in Figure 5b, it can be concluded that the reflectance between 500 and 700 nm is inversely proportional to the concentration of Chl-a, and the reflectance beyond 700 nm is basically insensitive to the concentration of Chl-a. From the spectra that are shown in Figure 5a, it can be concluded that the reflectance is proportional to the TSM concentration. The wavelength positions of the absorption valley at the wavelength of 660 ~ 680nm and the reflection peak near 700nm are very stable, and they do not change with the change of TSM concentration. Therefore, the absorption valley at 660–680 nm and the reflection peak around 700 nm can be used as characteristic bands for the inversion of Chl-a concentration.

From the reflectance spectra that are shown in Figure 5a, it can be concluded that almost the whole reflectance of the spectra are significantly and positively correlated with TSM concentration between 350 and 900 nm. From the reflectance spectra that are shown in Figure 5b, it can be concluded that the reflectances below 700 nm are highly sensitive to Chl-a concentration. The reflectances beyond 700 nm are almost completely insensitive to Chl-a concentration, and the reflectance after 830 nm is weak and noisy. Therefore, it can be concluded that the reflectance between 700 and 830 nm can be used for the inversion of TSM concentration.

#### *3.2. Inversion Model for Chlorophyll-a Concentration and Its Results*

Figure 6 shows the Pearson correlation between normalized remote sensing reflectance and Chl-a concentration. The correlation varied significantly for di fferent wavelengths.

**Figure 6.** Pearson correlation coe fficient (R) between normalized reflectance and Chl-a concentration.

Subsequently, we constructed isopotential maps of the linear correlation coe fficients of determination between the spectral indices and Chl-a concentration in the spectral interval 350–900 nm based on the original reflectance spectrum by establishing the spectral indices of reflectance di fference and reflectance ratio and using the least squares method to iteratively regress the spectral indices and Chl-a concentrations of Poyang Lake.

#### 3.2.1. Analytical Results for Spectral Data from August 2015

The coe fficients of determination between the ratio index and Chl-a concentration in Poyang Lake are generally low, as shown in Figure 7a. The coe fficients of determination for combination of wavelengths of 700–715 nm and 690–700 nm are relatively high, but they are still at a low level (highest <sup>R</sup>2=0.35). The coe fficients of determination of the di fference index for Chl-a concentration in Poyang Lake are higher than those of ratio index, as shown in Figure 7b. The combinations of spectral wavelengths that are sensitive to Chl-a concentration consist of the band near 650 nm and the band between 700 and 710 nm.

**Figure 7.** Linear correlation coefficients between the spectral index and Chl-a concentration. (**a**) August 2015 data (reflectance ratio). (**b**) August 2015 data (reflectance difference). (**c**) October 2015 data (reflectance ratio). (**d**) October 2015 data (reflectance difference). (**e**) January 2016 data (reflectance ratio). (**f**) January 2016 data (reflectance difference).

#### 3.2.2. Analytical Results for Spectral Data from October 2015

As shown in the isopotential map that was based on the original spectral data from October 2015 shown in Figure 7c, the coefficients of determination between the Chl-a concentration and ratio index in Poyang Lake are low. High coefficients of determination are observed for spectral combination of wavelengths of 680–690 and 690–700 nm, however the maximum value (R<sup>2</sup>=0.4) is still at a low level. The coefficients of determination between difference index and Chl-a concentration in Poyang Lake are higher than those between ratio index and Chl-a concentration, as shown in Figure 7d. The combinations of spectral wavelengths those are sensitive to Chl-a are 660–680 vs 720–730nm and 390–420 vs 870–900 nm.

#### 3.2.3. Analytical Results for Spectral Data from January 2016

The coe fficients of determination for the original spectral data from January 2016, which reached a maximum value of 0.5, were higher than the coe fficients of determination for the original spectral data from August and October 2015, as shown in Figure 7e. The spectral combinations with the highest coe fficients of determination are 440–470 nm vs 630–700 nm. The coe fficients of determination between di fference index and Chl-a concentration in Poyang Lake are higher than those of the ratio index, as shown in Figure 7f. The combinations of spectral wavelengths that are sensitive to Chl-a concentration are (1) 450–500 vs 500–540 nm, (2) 450–500 vs 650–700 nm, and (3) 360–420 nm and 760 nm. Table 3 summarizes the combinations of spectral wavelengths with the highest spectral index fit in Figure 7.


**Table 3.** Spectral response characteristics of Chl-a in Poyang Lake.

Note: R<sup>2</sup> is coe fficient of determination between spectral index and Chl-a concentration.

Overall, the results of the analysis show that the coe fficients of determination of the linear correlation between the ratio index and the Chl-a concentration were lower than the coe fficient of determination of the linear correlation between the di fference index and the Chl-a concentration. The di fference index corresponds to a wider range of sensitive bands than the ratio index. The spectral wavelengths that were sensitive to Chl-a concentration mainly corresponded to bands 1, 3, and 4 of the GF-1 image, and the bands of the GF-1 images have good overlap with the MODIS image. El-Alem et al. [35] have found that an APPEL (APProach by ELimination) model while using the combination of MODIS bands 1, 2, and 3 can be used to determine the Chl-a concentration in water. The maximum reflectance of Chl-a is in the near-infrared region. Furthermore, colored dissolved organic matter (CDOM), TSM, and backscattering also a ffect reflectance in the near-infrared band. CDOM has the maximal reflection in the blue band, so the influence of CDOM can be eliminated in the blue band. TSM is highly sensitive in the red band. Therefore, the red band can eliminate the influence of TSM on the reflectance spectrum of Chl-a. Pure water has strong absorption characteristics in the red and near-infrared bands, so the influence of backscattering can be eliminated.

The APPEL model was established based on the spectral characteristics of Chl-a, pure water, TSM, and CDOM, as follows [35]:

$$\text{APPEL} = \text{R(b}\_{\text{NIR}}) - \left[ (\text{R(b}\_{\text{BI}\,\text{UE})} - \text{R(b}\_{\text{NIR}})) \ast \text{R(b}\_{\text{NIR}}) + (\text{R(b}\_{\text{RED})} - \text{R(b}\_{\text{NIR}})) \right] \tag{4}$$

The combination of GF-1 bands 1, 3, and 4 was used to establish the APPEL model. Table 4 shows the details of the models for the inversion of Chl-a concentration in di fferent seasons.


**Table 4.** Details of the inversion models for the concentration of Chl-a.

Note: X is calculated by the APPEL (APProach by ELimination) model: X = R(B4) - [(R(B1) - R(B4)) \* R(B4) + (R(B3) - R(B4))], R(Bn) represents the reflectance of band n of the GF-1 image.

A summer inversion model for Chl-a concentration was established while using 33 sets of data that were measured in August 2015 (Figure 8a). Of the various models that were assessed, the inversion model that used quadratic polynomials had the highest fitting degree (R<sup>2</sup>=0.6936); this model was subsequently validated while using 10 sets of measured data (Figure 8b). The results showed that the RMSE of the model was 1.158 mg/m<sup>3</sup> and the MRPE was 3.99%. Additionally, an autumn inversion model for Chl-a concentration (Figure 8c) was established based on 25 sets of data measured in October 2015. Again, the model that used quadratic polynomials had the highest fitting degree (R<sup>2</sup>=0.6954). This model was validated with eight sets of measured data (Figure 8d). The results showed that the RMSE of the model was 0.90 mg/m<sup>3</sup> and the MRPE was 2.72%. Finally, a winter inversion model for Chl-a concentration was established while using 20 sets of data measured in January 2016 (Figure 8e). The model that used quadratic polynomials had the highest fitting degree (R<sup>2</sup>=0.6413). This model was validated using six sets of measured data (Figure 8f). The results showed that the RMSE was 0.44 mg/m<sup>3</sup> and the MRPE was 9.44%.

**Figure 8.** Test results for inversion models for Chl-a concentration for different periods: (**<sup>a</sup>**,**b**) calibration and validation results for August 2015; (**<sup>c</sup>**,**d**) calibration and validation results for October 2015; (**<sup>e</sup>**,**f**) calibration and validation results for January 2016. RMSE: root-mean-square error. MRPE: mean relative percentage error. R2: coefficient of determination.

The Chl-a concentration in the entire coverage of Poyang Lake was obtained while using the ENVI 5.3 software based on the results of the summer, autumn, and winter inversion models for Chl-a concentration (Table 3). Figure 9 shows the corresponding estimates of the spatial distribution of the Chl-a concentration in Poyang Lake for each of these three seasons.

**Figure 9.** The estimated Chl-a concentration in Poyang Lake during (**a**) August 2015, (**b**) October 2015, and (**c**) January 2016, obtained from GF-1 satellite images using polynomial inversion models.

Figure 9a shows the results of the Chl-a concentration inversion while using the GF-1 satellite image from August 2015. In the study area, August is a summer month and is also the flooding (wet) season with the highest water level. During this month, the water temperature of Poyang Lake rises, the water velocity is the slowest, and algal growth is rapid, which causes the Chl-a concentrations to be the highest of the year, i.e., 5–30 mg/m3. The highest concentrations of Chl-a are distributed in the waters near the shore of Poyang Lake and in the Nanji wetland national nature reserve in the south central of the lake.

Figure 9b shows the results of the Chl-a concentration inversion while using the GF-1 satellite image from October 2015. In October, which is an autumn month in the study area, the water level of Poyang Lake begins to decline and the water temperature to decrease. At this time, the estimated concentration of Chl-a in the lake decreased to between 2 and 15 mg/m3. The highest concentrations of Chl-a are distributed around the channel in the north of Poyang Lake mouth, which connects it to the Yangtze River, and in the main channel near the center of the lake.

Figure 9c shows the results of the Chl-a concentration inversion while using the GF-1 satellite image from January 2016. In the study area, January is a winter month and is also in dry season. During this month, algae grow slowly in Poyang Lake. The estimated concentration of Chl-a was generally low, ranging from 0–11 mg/m3, and the estimated distribution was more uniform than for August or October. The highest concentrations of Chl-a are distributed near the channel in the northern part of Poyang Lake, which connects it to the Yangtze River, and in the places where the Ganjiang, Fuhe, Xinjiang, Raohe, and Xiushui rivers flow into the lake.

#### *3.3. Retrieval Models and Results for Total Suspended Matter Concentration*

Figure 10 illustrates the Pearson correlations between normalized remote sensing reflectance and TSM concentration. The correlation coefficient shows obvious variation for different wavelengths.

**Figure 10.** The Pearson correlation (R) between normalized remote sensing reflectance and TSM concentration.

Two peaks and valleys were observed in the correlation curve for August 2015 (summer), as shown in Figure 10. The two peaks are located at wavelengths of 660~730 nm and 760~820 nm, respectively, and the maximum values of the correlation coefficients for the peaks are 0.85 and 0.75, respectively. The two valleys are located at wavelengths of 350~400 nm and 500~550 nm, respectively, and the maximum values of the correlation coefficients for the valleys are –0.76 and –0.72, respectively. For the August 2015 correlation curve, the maximum correlation coefficient (0.85) appears at a wavelength of 710 nm, and the minimum correlation coefficient (–0.76) appears at a wavelength of 371 nm. The correlation curve for October 2015 (autumn) shows different trends at wavelengths below and above 650 nm, respectively. Below 650 nm, the correlation coefficient is negative, reaching its minimum value of –0.96 at a wavelength of 590 nm; above 650 nm, the correlation coefficient is positive, reaching a maximum value of 0.98 at a wavelength of 756 nm. In January 2016, the water reflectance spectra that was most sensitive to TSM concentration, was observed at a wavelength of 723 nm, with the maximum correlation coefficient of 0.95. These sensitive wavelengths most corresponded to band 3 of the GF-1 image. Therefore, band 3 of the GF-1 image was used in the model for the retrieval of TSM concentration. We assessed the performance of five mathematical models for retrieval, which used linear, quadratic polynomial, exponential, logarithmic, and power-law equations, respectively, and selected the best-fitting model (i.e., the one with the highest coefficient of determination, R2) as the retrieval model. Table 5 describes the selected models for different seasons.

**Table 5.** Selected models for the retrieval of TSM concentration.


Note: Y represents the TSM concentration; X represents the reflectance of band 3 of the GF-1 image.

The field data that were measured in August 2015 were used to establish the summer retrieval model (Figure 11a). The retrieval model using a quadratic polynomial was found to have the best fit (R<sup>2</sup>=0.9003). This model was subsequently validated on 18 sets of measured data (Figure 11b). The results showed that the RMSE of the model was 6.96 mg/<sup>L</sup> and the MRPE was 27.52%. From the test results, it can be seen that the quadratic model in the single-band model predicted the summer TSM concentration of Poyang Lake well and the model had good stability.

**Figure 11.** Test results for the models for the retrieval of TSM concentration for three seasons: (**<sup>a</sup>**,**b**) calibration and validation results for August 2015; (**<sup>c</sup>**,**d**) calibration and validation results for October 2015; (**<sup>e</sup>**,**f**) calibration and validation results for January 2016.

The 17 sets of data that were measured in October 2015 were used to establish the autumn retrieval model (Figure 11c). The retrieval model using a power-law equation was found to have the best fit (R<sup>2</sup>=0.8614). The retrieval model was validated while using 16 sets of measured data (Figure 11d). The results showed that the RMSE of the model was 12.59 mg/<sup>L</sup> and the MRPE was 30.05%. It can be seen from the test results that the power-law model in the single-band model predicted the autumn TSM concentration of Poyang Lake well and the model had good stability.

The 14 sets of data that were measured in January 2016 were used to establish the winter retrieval model (Figure 11e). The retrieval model using the power-law equation was found to have the best fit (R<sup>2</sup>=0.6504). This retrieval model was validated while using 12 sets of measured data (Figure 11f). The results showed that the RMSE of the model was 5.37 mg/<sup>L</sup> and the MRPE was 20.83%. It can be seen from the test results that the power-law model in the single-band model predicted the winter TSM concentration of Poyang Lake well and the model had good stability.

The TSM concentrations for the whole of the Poyang Lake area were calculated while using the ENVI 5.3 software based on the results of the retrieval models for TSM concentration in summer (August 2015), autumn (October 2015), and winter (January 2016) (Table 4). Figure 12 shows a map showing the spatial distribution of TSM concentration in Poyang Lake.

**Figure 12.** The estimated spatial distribution of TSM concentration in Poyang Lake for three periods: (**a**) August 2015; (**b**) October 2015; and, (**c**) January 2016.

Figure 12a shows the result of the retrieval of TSM concentration that is based on GF-1 image from August 2015. The overall level of TSM concentration in Poyang Lake was relatively low in August 2015, and the lowest TSM concentrations occurred in the eastern, western, and southern parts of Poyang Lake. The concentration of TSM in the eastern part of the lake was generally below 100 mg/L, the concentration at the junction of the Xiu River and the Ganjiang River ranged from 0~68 mg/L, and the concentration in Junshan Lake (which lies to the south of Poyang Lake) ranged from 0~46 mg/L. The concentration of TSM was relatively high in the channel, which connects the north of the lake to the Yangtze River, and in the main channel in the center of the lake, due to the influence of sand mining in Poyang Lake [36]. In the northern part of the lake, the TSM concentration ranged from about 59–80 mg/L. The highest TSM concentration that was observed in the central channel was 103 mg/L.

Figure 12b illustrated that the overall TSM concentration in Poyang Lake in October was significantly higher than that in August. The highest concentration of TSM (254.43 mg/L) was observed in the channel connecting the northern part of the lake to the Yangtze River. This can be attributed to the fact that, in August and September, increased rainfall in the Yangtze River causes the water level of the Yangtze River near Poyang Lake to increase, which suppresses water outflow from Poyang Lake and causes the water from the Yangtze River to flow back into the lake. This flow causes the TSM concentration of Poyang Lake to reach its highest levels in the channel connecting it with the Yangtze River due to the high concentration of TSM in the Yangtze River. The increases in TSM concentration that were observed in other parts of Poyang Lake can be attributed to the continuous sand mining activity in the lake area.

Figure 12c illustrated that the TSM concentration of Poyang Lake varied between 0 and 201 mg/<sup>L</sup> in January, i.e., the maximum TSM concentration higher than that in August 2015. The highest TSM concentrations were observed in the channel that connects the north of the lake with the Yangtze River, and in the main channel in the center of the lake.
