5.1. Relative Value of Landsat 5 Optical and Quad Pol RADARSAT-2 SAR Data for Classifying Shoreline Types
Table 10 shows confusion matrices for three models constructed to demonstrate the relative value of Landsat 5 optical, and quad pol RADARSAT-2 SAR data for classifying shoreline types. The first model, based on all predictor variables, reached an overall independent accuracy of 93%. It is worth noting that this model was not significantly different from models generated by [
16] that included all the same inputs except the additional Landsat 5 variables generated for this analysis (i.e., only the spectral bands and NDVI values were used), nor the authors’ optimal model containing 14 Landsat 5, quad pol RADARSAT-2, and CDED variables. This is likely a result of many variables being highly correlated, as well as the high separability of classes.
For the second model, constructed with just Landsat 5 and CDED variables, overall independent accuracy is approximately 12% lower than the model that included the quad pol SAR predictor variables. This is largely due to increased confusion among substrates; a finding which is sensible since compared to optical sensors, the wavelengths at which SAR systems operate make them well suited for detecting differences in roughness, which tend to vary among the substrate classes. The roughness of the surface, measured relative to the wavelength of the SAR sensor, greatly impacts the amount of energy scattered back in the direction of the sensor, since this largely determines the degree to which reflection is specular or diffuse [
69]. Specifically, as roughness increases, reflection becomes more diffuse, increasing the amount of energy scattered back towards the sensor [
69]. Note that the RADARSAT-2 images evaluated in this research were especially suitable for detecting differences in roughness since they were acquired at a shallow incidence angle [
70]. In fact, findings by [
17,
18] provided a basis for selecting this beam mode as the authors observed improved separability amongst several class pairs, including several substrates, at shallow compared to steep angles.
Further, because spectral signatures are affected by the chemical composition of the surface, confusion amongst substrate classes for models containing only Landsat 5 data is in part due to different substrates being composed of the same rock type (e.g., Pebble/Cobble/Boulder and Bedrock composed of the same sedimentary rocks). To demonstrate this,
Figure 2 shows the spectral response of two features composed of the same material: one identified as Pebble/Cobble/Boulder, and the other as Bedrock. With the Landsat 5 image bands, many of the values for each class fall within a common range. However, with the quad pol RADARSAT-2 data, both exhibit distinct scattering behaviour (
Figure 2). For the Pebble/Cobble/Boulder class, both the Freeman–Durden double bounce and HV intensity values are higher, and fall outside the range of values observed for the Bedrock sample. Similar observations were made for other features throughout the study area. The coarse resolution of the Landsat imagery also likely played a role in the increased confusion amongst several classes. Recently, very promising results for sediment type discrimination based on very high-resolution Pleiades data have been observed [
71]. Given this, there is need for further research to better understand the effects of image resolution on the ability to differentiate substrate types.
By comparison, Random Forest models built with quad pol RADARSAT-2 and CDED data achieved better separability between many substrates, but also confused more vegetated and non-vegetated classes (e.g., Tundra versus Bedrock). Thus, independent overall accuracies were also lower (~13%) than models that included all predictor variables. This can similarly be explained by the fact that while vegetated and non-vegetated features typically absorb and reflect Near Infrared light differently, they can exhibit similar backscattering behavior.
Figure 2 shows that with select SAR variables, values for Tundra and Bedrock fall mostly within a common range due to the short stature and low-density of vegetation being mostly transparent at C-band, resulting in both surfaces exhibiting relatively similar surface roughness. Conversely, because healthy vegetation strongly reflects Near Infrared light, values for Tundra are much higher and fall outside the range of values observed for Bedrock. Note that Sand/Mud were also misclassified more times by models containing quad pol RADARSAT-2 and CDED data, which is also due to both classes exhibiting similar surface roughness.
These results clearly demonstrate the complementarity of optical and SAR data for shoreline mapping (especially in cases where only coarse resolution optical imagery is used), as both were required to achieve acceptable accuracies for all land cover types. The quad pol RADARSAT-2 data was more effective in discriminating several of the substrate classes, while the Landsat 5 imagery was preferred for separating vegetated and non-vegetated classes. With the Landsat 5 and CDED data alone, it was only possible to accurately discriminate Water, Bedrock, Wetland, and Tundra (uer’s and prducer’s accuracies >/=80% achieved), while with the quad pol RADARSAT-2 and CDED data, only Water, Pebble/Cobble/Boulder, and Wetland were accurately classified (use’s and proucer’s accuracies >/=80%).
These findings are consistent with [
16], who observed that both Landsat and RADARSAT-2 variables were among the most important inputs to their model. The authors of [
17] also observed increased confusion between several substrate types when classifying SPOT-4 spectral bands and NDVI values using pixel-based Maximum Likelihood. With the addition of RADARSAT-2 HH, HV and VV values however, user’s and producer’s accuracies increased for several classes, including Sand (by 38% and 12%, respectively) and Wood/Substrate Mix (by 10% and 12%, respectively). The authors of [
19] found that both quad pol RADARSAT-2 and SPOT-4 were useful in classifying multiple shoreline types using a hierarchical object-based classifier. With unsupervised SAR-based classifiers, the authors of [
18] could differentiate features with different roughness, though observed confusion between classes with similar roughness (e.g., Tundra vs. Mixed Sediment). SAR data therefore contribute positively to differentiating substrates and are useful in classifying shoreline types, which contributes to the increasing portfolio of remote sensing coastal observation methods [
72].
5.2. Comparing Performance of Random Forest Models Based on Quad Pol RADARSAT-2, Simulated Compact Polarized or Simulated Dual Polarized RCM Data in Combination with DEM Data
All models based on simulated RCM and CDED data achieved lower independent overall accuracies and were significantly different from the model based on quad pol RADARSAT-2 and CDED data (
Table 11). Results from analyses used to address the first objective of this research explain, in part, why there is greater confusion between classes when Landsat 5 spectral data are excluded from the model. On the other hand, the decrease in accuracy observed as a result of the substitution of quad pol RADARSAT-2 for simulated RCM data is mostly related to the decrease in information content of the latter [
24]. Conversely, the difference in NESZ values seems to have had less of an impact as indicated by the fact that user’s and producer’s accuracies did not decrease for classes with the lowest backscatter returns, including Water and Sand. This result is somewhat unexpected since at these incidence angles, values for these classes tend to fall close to or below both noise floors that were evaluated (i.e., below −19 dB for high, and below −25 dB for medium resolution data) [
17]. Further research is necessary to understand and verify these observations.
For some classes, the substitution of RADARSAT-2 for simulated RCM data had a negligible or varied impact on user’s and producer’s accuracies (e.g., producer’s accuracies for Mixed Sediment were higher in all cases, though user’s accuracies were generally lower). Conversely, for the wetland class, this resulted in a large decrease in user’s and producer’s accuracies in all cases (>/=6%, and up to 29%). This is mostly as a result of increased confusion with Tundra, for which user’s and producer’s accuracies were also generally lower for models constructed with simulated RCM data. These results are consistent with [
73] who similarly noted a decrease in the classification accuracy of wetlands when substituting quad pol RADARSAT-2 for simulated RCM data. Nevertheless, the authors found that the CP mode still achieved relatively high accuracies, and so suggested it was suitable for broad scale mapping. In [
74], the authors found that outputs from the Freeman–Durden decomposition applied to quad pol data were more effective in identifying flooded vegetation compared to the m-Chi decomposition applied to simulated RCM CP data.
It is worth noting that the Wetland class may have been more accurately classified if steep incidence angle data were used. The authors of [
17] observed that steep angle quad pol RADARSAT-2 was preferred for discriminating wetlands from tundra dominated by tall shrubs. This finding is sensible since, in theory, greater canopy penetration occurs at steeper angles resulting in greater sensitivity to sub-canopy conditions, including surface moisture and inundation [
75]. However, since shallow angle imagery are also preferred for roughness information, fusion of multi-angle SAR data may be necessary to achieve high accuracies for all classes. Given the four-day repeat pass cycle of RCM, multi-angle datasets will likely be more easily attained, thus will be a focus of future work.
At similar polarizations, models constructed with high or medium resolution mode data were not significantly different overall (based on McNemar’s statistic; 95% confidence interval), indicating that these two modes can be used interchangeably in some cases. It is notable however, that with the VV and VH polarization, user’s and producer’s accuracies for Wetland were substantially higher (10%) for models constructed with medium resolution mode data. Though neither achieved acceptable accuracies for this class, this does indicate that one mode may still be more suitable for specific applications.
For the same imaging mode, models constructed with simulated CP, HH and HV, VV and VH data were also not significantly different overall, again indicating that these polarizations can be used interchangeably for certain applications. For both high and medium resolution mode datasets, models constructed with HH and VV polarization data achieved significantly lower independent overall accuracies. This result is consistent with others that have demonstrated the value of HV over HH and VV for shoreline mapping [
16,
17].
5.3. Comparing Performance of Random Forest Models Based on Quad Pol RADARSAT-2, Simulated Compact Polarized or Simulated Dual Polarized RCM Data in Combination with Landsat 5 and DEM Data
With the exception of models containing medium resolution CP data, all others constructed with simulated RCM, Landsat 5, and CDED variables were significantly different, with lower overall independent accuracies, than the first model based on all quad pol RADARSAT-2, Landsat 5, and CDED data (
Table 12). However, by comparison, differences between these models (i.e., Model 1 versus Models 12–19;
Table 12) were less than differences between models constructed with quad pol RADARSAT-2 and CDED data, and simulated RCM and CDED data (i.e., Model 3 versus Models 4–11;
Table 11). Thus, in this research, the type of SAR data (i.e., quad, dual or CP) had less of an impact on overall accuracy when the Landsat 5 optical data was also included as an input.
As was observed for models based on SAR and CDED data only, for some classes the substitution of quad pol for simulated RCM data had only a slight or varied impact on user’s and producer’s accuracies. For Water and Bedrock, for example, differences between models containing quad pol RADARSAT-2 and simulated RCM data ranged from 0% to 3%. For Wetland and Tundra, differences were higher in some cases, ranging from 2% to 9%, and from 0% to 8%, respectively (
Table 12). Interestingly, all models constructed with simulated RCM, Landsat 5, and CDED data achieved accuracies that were considered to be within an acceptable range for operational mapping (i.e., >/=~80%, with the only exception being the user’s accuracy for Mixed Sediment which was 79% when classified with the Simulated RCM high resolution HH and VV imagery, Landsat 5, and CDED data). As such, it is expected that these data, which will be available at a greater temporal frequency and wider swath width than RADARSAT-2, will complement current efforts focused at mapping shorelines throughout the Canadian Arctic [
16,
17,
18].
Note that these results are consistent with [
76] who used Random Forests to classify Peatlands in Southern Ontario. The authors similarly observed that when classified in combination with Landsat 8 optical and Shuttle RADAR Topography Mission DEM data, there was not a significant difference between models that contained quad pol RADARSAT-2 or simulated RCM data.
With the exception of the VH and VV polarization, models constructed with high or medium resolution mode data were not significantly different at similar polarizations, again demonstrating that in some cases both imaging modes can be used interchangeably. In fact, the maximum difference in user’s and producer’s accuracies between models containing high and medium resolution data was 5%, and independent overall accuracies only differed by 1%. For VH and VV polarization however, a statistically significant difference was observed between models constructed with high and medium resolution mode data, with the latter achieving higher accuracies for Sand/Mud and Mixed Sediment.
Similarly, with similar imaging modes, models constructed with simulated CP, HH and HV, and VV and VH data were not significantly different overall. Notably though, user’s and producer’s accuracies for Wetland were highest with CP data. Since these features typically represent important species habitat and are sensitive to the effects of climate change and of oiling, it is essential that they are accurately classified. This justifies preference for this beam mode for certain applications, including shoreline mapping. As was observed when Landsat 5 data were excluded from models (
Table 11), those constructed with simulated HH and VV data achieved significantly lower overall accuracies compared to models containing other data for other polarizations, which is again consistent with observations by others that HV is generally preferred over HH and VV for shoreline mapping [
16,
17].
5.4. Determining the Extent to which Model Data Load Can Be Reduced without Impacting or Possibly Improving Overall Accuracy
Given the number of datasets evaluated in this research, the decision was made to select one to evaluate the effect of reducing the model data load. Given that [
16] already evaluated the effect of reducing the dimensionality of the quad pol RADARSAT-2, Landsat 5, and CDED dataset, we chose to evaluate the Simulated RCM medium resolution mode CP, Landsat 5, and CDED dataset (i.e., Model 19). This dataset was also evaluated because it achieved the highest accuracy of all models containing simulated RCM data, performing most similarly to the model containing quad pol RADARSAT-2, Landsat 5, and CDED data.
Results from this analysis indicate that multiple methods can be effective in reducing the number of inputs to Random Forests models, without affecting overall accuracy (
Table 13). In this research, performance of the model did not vary significantly based on the reduction method, thus demonstrating, as others have observed [
16], that Random Forests is not highly sensitive to the type and number of inputs. We expect that this is especially the case here given that many variables were highly correlated, and classes were highly separable.
It is worth noting that, for the first method, a threshold value of
r > 0.5 was found to be effective in reducing redundant information without affecting classifier accuracy. In addition, the second method, while not taking into account the possible spreading of importance values among correlated inputs [
64], still achieved the same accuracy, while also being the most efficient approach (in terms of computation expense and required user intervention). These results also show that the CDED data was not needed for accurate discrimination of the land covers evaluated in this research (
Table 13). We suspect that this is due to a combination of the DEM being provided at a coarse resolution, and the fact that most features in the study area are relatively low lying and flat.