Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
Joplin held out (wind only damage classifier) 90% of non-Joplin Joplin only 0.74 0.89 0.42 0.54 0.77 0.60
Nepal held out 90% of non-Nepal Nepal only 0.63 0.42 0.17 0.23 0.02 0.06
Nepal held out (flood only damage classifier) 90% of non-Nepal Nepal only 0.64 0.54 0.12 0.27 0.07 0.14
TABLE III
PIXEL -WISE F1SCORE ACROSS VARIOUS SPLITS OF THE X BD DATASET . W E TEST GENERALIZATION PERFORMANCE OF MODELS ON THE JOPLIN WIND
AND NEPAL FLOODING EVENTS IN TWO SETTINGS :ONE IN WHICH WE TRAIN ON allAVAILABLE DATA THAT IS NOT FROM THE SPECIFIC EVENT ,AND
ANOTHER SETTING IN WHICH WE TRAIN THE DAMAGE CLASSIFICATION DECODER ON OTHER WIND -ONLY EVENTS (FOR TESTING ON THE JOPLIN
EVENT )AND OTHER FLOOD ONLY EVENTS (FOR TESTING ON THE NEPAL FLOODING EVENT ).
•The same Python virtual environment for all experiments
to remove the effect of different packages on the
performance.
Additionally, the inference times reported in Table IV
include the file I/O, model loading, pre-processing, and post-
processing costs associated with each approach and therefore
represent an upper bound on the time taken to process any
given 1024×1024 input (i.e. when running such approaches
over large amounts of input, the models would only need to
be loaded from the disk a single time).
As previously discussed, the xView2 challenge2encouraged
participants to optimize for leaderboard performance instead
of throughput. As such, many of the top-placed solutions used
techniques such as ensembling and test time augmentation, as
well as larger, more complex models in order to improve their
performance at the cost of inference speed. The top-performing
solution, for instance, consists of an ensemble of 12 models
that are run 4 times for each input (test time augmentation with
4 rotations). These solutions are prohibitively costly to run
on large inputs. For example, the Maxar Open Data program
released ∼20,000km2of pre- and post-disaster imagery
covering areas impacted by Hurricane Ida in 2021. Assuming
the inference times from Table IV, 0.3m/px spatial resolution
of the input imagery and $0.9/hr cost of running a Tesla K80
(based on current Azure pricing), the first place solution would
cost $6,500 to run, while our solution would only cost $100
to run. In this case, our solution would generate results for the
area affected by Hurricane Ida in 4.7 days while the first place
solution would take up to 301.4 days using a single NVIDIA
Tesla K80 GPU.
Finally, we benchmark our proposed solution in an
optimized setting compared to the above setting: we load data
with a parallel data-loader (vs. loading a single tile on the main
thread), we run pre- and post-processing steps on the GPU,
we maximize the amount of imagery that is run through the
model at once (vs. running on a single 1024×1024 tile of
imagery), and we use the most recent version of all relevant
packages (vs. the earliest version pinned in the environments
from the xView2 solution repositories). Here, we find that
our model is able to process 612.29 square kilometers per
hour compared to 89.35 square kilometers per hour under the
2Most machine learning competitions follow a similar format, whereby
participant solutions are only ranked in terms of the held-out test set
performance.same assumptions in the previous setup despite using the same
hardware. In this case, our model could process the Hurricane
Ida imagery in 2 days at a cost of $14.7. The stakeholder’s
baseline solution’s speed is 1000 square kilometers per hour
on Azure NC12 GPU. We project our runtime to be 20% faster
than their baseline solution on a similar GPU.
Method Inference time (s) sq. km/hr
xView2 1st place 245.75 (0.73) 1.38
xView2 2nd place 121.03 (0.36) 2.81
xView2 3rd place 108.21 (0.6) 3.14
xView2 4th place not reproducible not reproducible
xView2 5th place 10.94 (0.06) 31.07
Our method 3.8 (0.02) 89.35
TABLE IV
COMPARISON OF BUILDING DAMAGE MODEL INFERENCE TIMES ON A
SINGLE 1024 X1024 PIXEL TILE FOR DIFFERENT METHODS USING A
SINGLE TESLA K80 GPU ( ON AN AZURE NC6 MACHINE ). T IMES ARE IN
SECONDS AND ARE AVERAGED OVER THREE RUNS WITH A STANDARD
DEVIATION IN PARENTHESES . THE RESULTS FOR THE WINNING X VIEW2
SOLUTIONS ARE REPRODUCED THROUGH THE OFFICIAL GITHUB
REPOSITORIES PUBLISHED FOR EACH ,WHERE THE ONLY MODIFICATIONS
TO THE ORIGINAL CODE WAS TO ENABLE GPU PROCESSING FOR EACH
INFERENCE SCRIPT . THE RIGHTMOST COLUMN SHOWS THE INFERENCE
SPEED IN TERMS OF (SQ.KM)/HR ASSUMING A 0.3M/PIXEL INPUT SPATIAL
RESOLUTION .
VIII. W EBVISUALIZER TOOL
In contrast to standard vision applications, semantic
segmentation models that operate over satellite imagery need
to be applied over arbitrarily large scenes at inference-time. As
such, distributing the imagery and predictions made by such
models is non-trivial. First, high-resolution satellite imagery
scenes can be many gigabytes in size, difficult to visualize
(e.g. requiring GIS software and normalization steps), and may
require pre-processing to correctly align temporal samples.
Second, the predictions from a building damage model are
strongly coupled to the imagery itself. In other words,
only distributing georeferenced polygons of where damaged
buildings are predicted to be is not useful in a disaster response
setting. The corresponding imagery is necessary to interpret
and perform quality assessment on the predictions.
Considering these difficulties, we implement a web-based
visualizer to distribute the predictions made by our model over
satellite image scenes. This approach bypasses the need for any
specialized GIS software, allowing any modern web-browser
Fig. 6. Screenshot of the building damage visualizer instance for the August,
2021 Haiti Earthquakes. The left side of the map interface shows the pre-
disaster imagery while the right side shows the post-disaster imagery. The
slider in the middle of the interface allows a user to switch between the pre-
and post-disaster layers to quickly see the difference in the imagery. Finally,