October 2019 Layout Flipping Full - page 771

areas (region B, C), and a special mismatching area (region D)
which contained both building area and nonbuilding area.
Firstly, it seemed that the SRMD-DSMs were always the closest
results to the corresponding Ref-DSMs, even in the mismatched
region, which indicated that SRMD could achieve the best per-
formance among these image reconstruction models no matter
which type the terrain belongs to. The conclusion was the same
to the image reconstructed results shown in Table 1 and Figure
3. This indicated that the generated DSM would be better with
a better SR model, so the problem of DSM quality improvement
might be converted into an easier image SR problem.
Secondly, in the building region (region A), the quality
difference among these DSMs were slight and nearly invis-
ible. It was explainable as the resolution of used images were
high enough (0.5 m), so enough accurate feature points in the
building edge could be obtained for subsequent DSM genera-
tion. Even when most of this region seemed to be the same,
DSM upscale and bicubic methods failed to reconstruct the
complete building in the red box, while all CNN-based SR
methods could obtain the whole building edge, and the result
of SRMD was nearly same to the reference. On the contrary, the
visual difference was more distinct in nonbuilding regions
(region B, C) among the SR-DSMs. For example, in the red box
of region B and C (both belong to vegetation covered areas),
BI-DSMs performed the worst and even could not reconstruct
the correct terrain, while all the SR-DSMs could, more or less.
And compared to the DSM upscale method, CNN-based SR
models were more robust and less likely to produce incorrect
results. It was because more terrain features and mapping rela-
tionships were learned from the samples when training these
models, thus the reconstructed images could represent the
real terrain better. All the results confirmed the significance
of enhancing image texture for better DSMs generation, and
further verified the feasibility of the proposed application.
Objectively, as illustrated in Table 2, in nonbuilding re-
gions, CNN-based SR models always outperformed the bicubic
upscale method, especially in some complex terrains, e.g. the
red box area of region C. As shown in Table 2, the RMSE and
MRE of this region were even up to 118.8005 and 17.83% in BI-
DSM, while the SR-DSMs had lower RMSE and MRE. In particular,
the SRMD-DSM reached the lowest RMSE and MRE, which was
just about 2/3 and 1/3 of the BI-DSM. However, even though
the experimental results clearly verified that the SRMD method
could obtain a much better DSM than the direct upscale meth-
od, it must be noticed that the RMSE results of region B and
region C were still too high. As the abnormal RMSE occurred
with vegetation covered regions (region B, C), we constructed
further experiments to investigate the specific reason. Partial
experiment results are shown in Figure 5 and Table 3.
Reference DSM UPSCALE Bicubic
SRCNN VDSR SRMD
Region D: mismatching area (the red box: mismatch areas)
Figure 4. 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).
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING October 2019 771