October 2019 Layout Flipping Full - page 770

“SRCNN-DSM”, “VDSR-DSM”, and “ SRMD-DSM ”. And all the DSMs
obtained from CNN-based SR models were denoted as “SR-DSMs”.
Generally, it was obvious that the CNN-based methods
could generate better subpixel level DSMs in any terrain situ-
ations compared to directly DSM upscale from Table 2. And
the results in Figure 4 also showed that the DSMs obtained
from the proposed method had higher fidelity and contained
more terrain details. Besides, as the SR-DSMs were obtained
from higher-resolution image pairs, there were less mismatch
compared to the directly DSM upscale results.
For the sake of comprehensiveness, both subjective and objec-
tive evaluations were taken into consideration for analysis. Sub-
jectively, in Figure 4, the selected four regions could be mainly
divided into three parts: building areas (region A), nonbuilding
Reference DSM UPSCALE Bicubic
SRCNN VDSR SRMD
Region B: mountain area (the red box: forested areas)
Reference DSM UPSCALE Bicubic
SRCNN VDSR SRMD
Region C: a main road area (the red box: forested areas)
Figure 4 continued. Comparison of generated DSMs from different methods in five regions of several different types of terrain.
(Region A is a building region, region B is a mountain region, region C is a main road region, and region D is a special region
which contained mismatch). Concluded on next page.
770 October 2019 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING