3.1. Landcover Classification
The segmentation of the WorldView-3 imagery yielded nearly 86,000 objects, each of which was then classified as belonging to one of the six land cover classes (
Table 5). These segmentation and classification results were subsequently used for delineation and parameterization of the subcatchments in the hydrologic model. Segmented and classified images are shown alongside the original pansharpened Worldview-3 image for a magnified sample area (
Figure 4). In general the segmentation parameters adopted from Qian et al. [
23] were found to generate appropriately sized segments for the current application. Initial segmentation trials (i.e., those not following Qian et al. [
23]) were found to often generate objects which were smaller than desired for the current application, e.g., individual cars on roads were delineated as separate objects.
Classification results for the whole study area (
Figure 5) showed that impervious areas (i.e., buildings, PPPA, roads) and green areas (i.e., trees, grass) account for approximately 52% and 46% of land cover, respectively. The trees class was the largest, covering nearly a third of the study area. The total tree area is likely slightly overestimated considering that tree canopy may extend beyond the pervious area where the trees are planted in some cases. Green areas are scattered throughout the whole study area, however are much more concentrated in the areas of Fragrant Hills and Old Summer Palace in the west and the Olympic Park in the east (
Figure 5).
3.2. Classification Accuracy Assessment
The Confusion matrices for classification based on the Bayes classifier alone (
Table 6) and the Bayes classifier with the refining rules (
Table 7) quantify the accuracies of each approach. In the matrices, “overall accuracy” refers to how many of the reference sites were classified correctly relative to the total number of reference sites. Producer’s accuracy represents how often real features on the ground are correct on the classified map, while the user’s accuracy represents how often the class on the map will actually be present on the ground. The Kappa coefficient evaluates how well the classification performed, where a value of 0 indicates that the classification is equivalent to random classification, and a value close to 1 indicates that the classification is significantly better than random.
The accuracy assessment indicated that when the refining rules were applied to the initial classification produced by the Bayes classifier, both the overall accuracy and the Kappa coefficient increased significantly, from 63% to 76%, and from 0.56 to 0.72, respectively. While there is no broad consensus on an overall accuracy that should be achieved for an application like the current one, a commonly referenced standard in the broader remote sensing literature suggests 85 to 90% [
34]. The current study fell short of this target both before and after application of the refining rules. The required Kappa coefficient also depends greatly on the specific application, however the values achieved in the current study fall in the category of “substantial agreement” (i.e., from 0.61 to 0.80) based on the Kappa statistics groupings suggested by Landis and Koch [
35].
The greatest improvement after the refining rules were applied was the user’s accuracies for the Roads and PPPA classes (i.e., from 45 to 81% and from 14 to 68%, respectively). The spectral signatures of these mostly pavement covered areas tended to be very similar, and therefore the supplementary topological data (i.e., the right of way layer) was required to more consistently differentiate between them. Another observed significant improvement was the producer’s accuracy for the buildings class, which increased from 48 to 92% due to the refining rule which considered the difference in values between the DSM and DTM. The accuracy assessment results have thus demonstrated the value of supplementary data sets when performing GEOBIA classification of VHR imagery in urban areas with a high concentration of impervious surfaces with similar spectral signatures. Classification of green areas (i.e., trees and grass) was also improved with the refining rules. Using the Bayes classifier alone, tree objects were frequently (29 times out 100) incorrectly classified as grass. The refining rule attempted to improve this through the use of a threshold value of the standard deviation of the NIR band (
Section 2.3.3), resulting in an increase in the producer’s accuracy of trees from 67 to 80%.
However, even after the refining rules were applied not all of the individual accuracy metrics were significantly improved, and a few had values somewhat lower than before the refining rules were applied. The buildings class user’s accuracy, for example, lowered after application of the refining rules. Most objects that the Bayes classifier alone classified as buildings were in fact buildings, however there were also many (52 out of 100) building objects not classified as buildings. This high user’s accuracy combined with low producer’s accuracy represents an error of omission (i.e., not enough pixels have been classified as a given class). It should be noted that after application of the refining rules the Building class producer’s accuracy increased much more than the user’s accuracies decreased, considered an acceptable tradeoff that contributed to a higher overall accuracy. After application of the refining rules, 92 of the 100 building objects were correctly identified as Buildings.
The Kappa coefficients and overall accuracy obtained for the classification based on the Bayes classifier alone (i.e., 0.56 and 63%, respectively), were not nearly as high as the 0.96 and 96% obtained by Qian et al. [
23], which also used 125 training samples and the Bayes classifier for GEOBIA in an overlapping study area in Beijing. The accuracies found in the current study were also lower than the mean overall accuracy of 83.6% given for Worldview data based classification studies in general [
19]. There are likely a number of explanations for these discrepancies. Firstly, the classes used by Qian et al. [
23] did not differentiate between trees and grass or between the various types of impervious cover and these are by far the most common misclassifications in the current study. If the six classes chosen for this study were reduced to three as in Qian et al. [
23] (i.e., water, green areas and impervious areas), the overall accuracy would exceed 90% and be in line with the results found in that study. In general, the more specific and numerous the classes are, the more difficult it is to correctly assign objects to the correct class. This is confirmed by the review by Ma et al. [
19], where it was found that there is a negative correlation between the overall classification accuracy and the number of classes defined.
Secondly, the fact that the current study area was relatively large with likely higher variation within each of the land cover types may account for the somewhat lower accuracies. Applying GEOBIA to a smaller area may yield higher accuracies as urban features tend to be more similar to nearby features than they are to features in other parts of the city. For example, most of the residential buildings in one neighborhood of a city may have similar spectral characteristics because of the construction materials and methods used when that area was developed.
Fortunately, the most common misclassifications were not highly impactful in terms of overall stormwater runoff properties. For example, misclassifying roads as PPPA and vice versa does not significantly impact the results of the hydrological modelling for this particular study as both classes were assumed to have the same runoff properties and rates of conversion to permeable pavement in the LID scenarios. Misclassification of roads into the building class, for example, would be more impactful to the hydrologic modelling results because each of these class’ respective LIDs (i.e., permeable pavement and green roofs) perform differently.
3.3. Hydrologic Modelling
The land cover classification obtained from the GEOBIA procedure presented above aided in the development of a hydrologic model that returns realistic results [
28] and can be run reasonably efficiently (i.e., 1 h per scenario, per event, using a 3.4 GHz i7 processor and 32 GB RAM). Runoff hydrographs as computed by the SWMM model (
Figure 6) show that runoff peak flows and volumes in the Qing River decrease as the level of LID implementation increases for all return periods, as is expected. Also as expected, the computed runoff volume reductions relative to the Baseline scenario runoff volume (
Table 8) increases with an increasing level of LID implementation and decreases with an increase in return period of the event. However, the volume reduction for the Low LID scenario did not change significantly with return period, i.e., 62 and 61% of runoff volume was retained for the 3 and 100 year events, respectively. In the High LID scenario the volume reduction changes somewhat more with return period (i.e., 82 and 77% for the 3 and 100 year storms, respectively). This is likely because the extent of LID coverage limits the amount of runoff that can be captured for lower return period events, whereas the storage capacity of the LID is more of a limiting factor for higher return period storms (i.e., many LID are overflowing during large storms). There is therefore a need to know where the rainfall can be captured within the city in addition to the volume that can be captured at that location. Use of the detailed land cover data allows for these spatial considerations in hydrologic modelling. Although there is likely some potential for controlled conveyance of water to LID offsite, the general principles of LID call for onsite capture, and this is what was represented in the SWMM model. That is, runoff can only drain to LID within the subcatchment, or in the case of rain gardens, also from the adjacent impervious Road, Building or PPPA subcatchments.
While the 77% volume reduction of the 100 year event for the High LID scenario represents a substantial decrease in runoff, it should be noted that this scenario represents an ambitious level of implementation that will require a large portion of the study area to be retrofitted from its current state. Nevertheless, the results indicate that it is physically possible for a significant portion of a large event to be captured using LID infrastructure built to common specifications and considering the limitations of the urban landscape.
It should be noted that the SWMM model currently does not include representation of overland flow routes. Runoff that cannot be accommodated within the sewer system simply ponds on top of the model node until there is capacity available in the storm sewer system. For this reason, peak flows of very large events may not be modelled accurately if significant overland flow would occur in reality. However, even for the larger storms the total volume reductions provided by LID systems are likely accurate as the storage within the LID will be filled before overland flow occurs in the model. To more accurately represent peak flow reduction, lag times, or surface ponding depths for high return period storms, the current 1D model would need to be coupled with an additional 1D or 2D hydraulic model to simulate surface flow paths. If the current model were at some point to be enhanced with such representation of overland flow, the detailed land cover map developed in this study will provide further value in terms of identifying the overland flow obstacles (i.e., buildings) and assigning roughness values to the various urban surfaces.