3.1. Correlation Analysis between Single Parameters and Fuel Loads
Correlation analysis was executed between the satellite data and FLs in this section, as shown in Figure 2
For SFL, HV polarization of the L-band has the highest correlation, followed by HH polarization. Compared with SAR data, optical data generally has a lower correlation, with an average R2 of about 0.4. SWIR1 has the highest correlation among the optical bands.
For BFL, the correlation trend of each parameter is similar to SFL, HV polarization of the L-band is the most sensitive variable followed by HH polarization. Among optical bands, Green and SWIR1 bands have the highest correlation, while the correlation of the optical band with BFL is higher than that of SFL.
For FFL, both HV and HH polarizations can indicate it well. All optical bands are more correlated with FFL than SFL or BFL, especially the Green band. The SWIR1 and SWIR2 bands also have a high correlation with FFL, producing an R2 of about 0.6.
Overall, the HV polarization of L-band SAR data has the highest correlation with three types of forest fuel loads, and there is almost no correlation difference between different FLs. As for optical data, most bands have the highest correlation with FFL, followed by BFL, and the lowest for SFL. Specifically, the Green band has a poor correlation with SFL but a relatively high correlation with BFL and FFL. The results are reasonable since the optical signal mainly comes from the foliage in the forest canopy.
More detailed information on the relationships between satellite data and FLs is shown in Figure 3
. It can be found from Figure 3
a that both HV and HH polarizations increase with increasing SFL. The HV polarization has a strong positive correlation with SFL where the R2
is 0.75, while the HH backscattering coefficient has a relatively weaker positive relationship with SFL. In contrast, optical data have a weak negative correlation with SFL and the correlation is not as good as SAR data. Figure 3
b shows scatter plots between parameters and BFL. Similar to SFL, HV shows a strong positive relationship with BFL, while optical data shows a relatively weaker negative relationship with BFL. A similar correlation trend can also be found in Figure 3
c. However, we found optical data and HH polarization have a higher correlation with FFL than that of BFL and SFL. Spectral information can express FFL as more representative than SFL and BFL. These analyses may indicate that optical data are more suitable for FFL estimation but less so for SFL or BFL.
3.2. Fuel Load Estimation from Satellite Data
Through the correlation analysis between the satellite data and FL in the previous section, it is found that the highest correlation with each kind of fuel load is HV polarization. Additionally, the relationship of optical data to FFL, BFL and SFL is weakened in turn. Figure 4
shows the results of further study on the capability of satellite data to estimate various FLs using optical bands, SAR data and both of them, respectively. The R2
in Figure 4
is the optimal result of the combination situation of each number parameter in Table 2
. Take the SFL estimation using both optical and SAR data as an example, the R2
is 0.76 when the number of parameters is three, indicating that the best result was achieved when combining HV, NIR and SWIR2 variables.
For SFL, the estimation effect of all optical bands is not good; the average R2 is 0.36, among which SWIR1, NIR and Blue bands are the variables with the highest frequency, while SAR has a good performance on SFL, and the prediction accuracy R2 of a single HV polarization reaches 0.71. When combining two kinds of satellite data, as the number of parameters increases, the R2 slightly increases especially when NIR and SWIR2 are added based on HV, but then it begins to decrease. Compared to the individual SAR data (HV), the combination of optical and SAR data has little effect, with the R2 only increased by 0.04. This indicates that optical data cannot provide more information to characterize SFL than SAR data, and the introduction of more optical parameters will bring more errors and uncertainties, resulting in a decrease in model accuracy. Therefore, it is recommended to use only SAR data or introduce a few optical variables to carry out SFL estimation.
For the BFL, the performance of optical data is better than SFL. When selecting Green, Red and SWIR1 bands, the R2 reaches 0.56, and the Green band appears in every optimal model. The performance of SAR data is slightly inferior to that of SFL, but it can also characterize BFL well, especially HV polarization (R2 = 0.70). As for the combination of them, the trend of the R2 is similar to that of SFL; the accuracy will not always increase with the increase in variables but tends to slightly decrease. When a model is composed of six variables or more, the performance of predicting BFL begins to decrease since more variables may introduce uncertainty to the estimation. Compared to individual SAR data (HV), the combination of optical and SAR data (i.e., HV, Green and red) has a positive effect, with the R2 increased by about 0.1. This indicates that optical data also contributes to the estimation of BFL. Therefore, the BFL estimation can mainly focus on SAR data and the combination appropriate spectral bands.
For FFL, the optical spectral bands perform best with an R2 of 0.66 when Blue, Green and Red bands are combined. The Green band is the best indicator of FFL among the six bands, followed by Blue. Additionally, SAR can predict FFL well individually. Besides, the addition of optical data to SAR has significantly improved the model performance (R2 = 0.82) and outperformed any single data. The biggest contribution to the FFL estimate comes from HV, Green, HH and Red. However, after the number of variables is greater than four, the prediction accuracy tends to be stable and no longer increases. This is because that combination already contains the main information required by the FFL which not only has the forest structure information reflected by HV and HH, but also the spectral reflection information reflected by the optical band. Therefore, the FFL estimation should combine optical (especially the visible bands) and SAR (especially the HV polarization) data.
Overall, SAR data performed better for three types of FLs than optical, which is consistent with the correlation analysis in the previous section, while the combination of the two kinds of satellite data is better than either data alone. It is worth noting that the prediction accuracy of SFL had a clear downward trend with the increase in variables, while BFL and FFL had no such phenomenon. Besides, it was found that the most sensitive variables for the three types of FLs are different, indicating it is necessary to estimate them separately according to fuel load features (e.g., vertical structure in the forest) and different imaging mechanisms.
The relationships between variables explored in this study are shown in Table 3
higher than 0.75 displayed in bold font. It is found that the closer the wavelength is, the higher the correlation between variables, such as SWIR1 and SWIR2, Green and Red. In contrast, the correlation between optical band data and SAR polarization data is relatively low. Specifically, the correlation between HV and the data of various optical bands is relatively low, especially with NIR. This may explain the phenomenon that the combination of HV and optical band can improve the FL prediction since they can carry more information on the FL than any single data. However, the correlation between different variables is generally high, probably because the study sites in this experiment are stand scale which reduces the regional heterogeneity of the remote sensing data. In the final prediction FL model selection, we select as few variables as possible (less than three) while considering the model performance to make the model more concise and avoid the introduction of redundant errors.
Based on the analyses above, we comprehensively considered the model performance and correlation between variables. A total of nine models including Opt, SAR and SAR + Opt prediction models for SFL, BFL and FFL are selected. Each selected optimal model corresponds to no more than three parameters (i.e., FLs estimated by optical data alone choose the best model composed of three parameters; FLs estimated by SAR data alone choose the model composed of two parameters; FLs estimated by both optic and SAR data choose the best model composed of three parameters), and the corresponding parameters are all contributing the most. The detailed performance of these models is shown in Table 4
and Figure 5
Single optical data has the best prediction effect on FFL (R2 = 0.66) and the worst performance on SFL prediction (R2 = 0.37), while single SAR data performs similarly to SFL, BFL and FFL predictions (R2 = 0.72, 0.70, 0.72). Compared with the estimation with single SAR data, the addition of optical data has significantly improved the estimation accuracy for BFL and FFL (R2 increased by 0.1 and 0.07), but slightly for SFL (R2 increased by 0.04). Moreover, integrated L-band SAR and optical data can estimate three types of FL better than any single data (R2 = 0.76, 0.80, 0.79), indicating that the combination of optical spectral information (especially the Green, Red and SWIR2 bands), and SAR data polarization information (especially HV polarization), can describe the forest fuel load well.