text
stringlengths
0
820
the building damage predictions are shown as polygons with varying shades
of red corresponding to increasing damage. The visibility of these predictions
can be toggled in the interface so that a user can see the underlying imagery.
to view the imagery and predictions, and doesn’t require users
to have any formal GIS experience as all imagery is pre-
rendered. Specifically, users can:
1) Toggle back and forth between the pre- and post-disaster
imagery to easily see the differences;
2) Change the visibility of the damage predictions to see
the extent of the damage;
3) Show standard layers (e.g. OpenStreetMap or Esri World
Imagery) for additional spatial context.
This is implemented with open-source tools including:
GDAL3, leaflet4, and Docker5.
An instance of our visualizer is shown in Figure 6 for
a scene from Jeremie, Haiti after the Haiti Earthquake in
August, 2021. The tower of the Cathedral of Saint Louis Roi
of France (middle of the scene) is classified as damaged by the
model and can be seen to be destroyed. The code for running
inference with our final building damage model, as well as
setting up an instance of the building damage visualizer tool
is publicly available6.
Fig. 7. Full screenshots of pre- and post-disaster images shown partially in
the building damage visualizer instance in Figure 6 for better visibility. 2021
Haiti Earthquakes.
3https://gdal.org/programs/index.html
4https://leafletjs.com/
5https://www.docker.com/
6https://github.com/microsoft/Nonprofits/IX. D EPLOYMENT AND APPLICATIONS
Automating building detection and damage assessment has
the potential to tremendously speed up disaster response
processes [6], [27], which are of critical importance for
humanitarian organizations to estimate the geographical extent
and the severity of a disaster and plan accordingly [28]. To
ensure that such an assessment can be delivered in time,
this study’s stakeholder is implementing the proposed model
within a scalable, distributed computing system. Within this
system, satellite images are divided among many identical
instances of the model, which process them in parallel. This
guarantees a fixed computation time with any number and
size of input satellite images. The model’s output is then
shared with the wider humanitarian network in three ways: the
aforementioned web visualizer, the open data-sharing platform
“Humanitarian Data Exchange”, and via man-made maps
(in digital or printed format), which can be directly sent
to and used by first responders in the field. This ensures
the rapid diffusion of information among all stakeholders
involved in disaster response management. It is worth noting
that our stakeholder’s experience with applying such tools in
humanitarian settings and discussions with practitioners have
highlighted the importance of two aspects. First, the value
of automating damage assessment lies in speed, rather than
accuracy. Regardless of visible damages, detailed ground-level
inspections by trained personnel are still needed to assess
the structural integrity of a building [29] and it is unlikely
that remote sensing technology will replace that in the near
future. For this reason, the focus of satellite-based damage
assessments should be to provide broad numerical estimates
as fast as possible, rather than building-level prescriptions.
Secondly, while the immediate response is primarily informed
by disaster impact (which can be quantified by the number of
damaged buildings, among other metrics), long-term shelter
recovery programs must take into account several other
contextual factors, such as the socio-economic conditions of
affected people and land ownership [30]. Because of this,
information on building damage often needs to be combined
with other data to be useful. Providing raw data including
geo-referenced building footprint masks and corresponding
damage levels to the humanitarian community is necessary to
enable these analyses. Furthermore, as we discussed in section
VI, trained models based on the proposed approach might
not be robust to significant distribution shift across different
geographies; as such, domain adaptation techniques need to be
explored to address the data biases issues [31]. In this context,
active learning and human-machine collaboration approaches
have been discussed in [32] and [33].
X. C ONCLUSION
Natural disasters’ frequency is growing; thus, the impact
of such events on communities continues to increase. The
strategic response of humanitarian organizations to allocate
resources and save lives after disasters can be improved by
using AI tools. We propose a convolutional neural network
model that uses satellite images from before and after natural
disasters to localize buildings using the UNet model and
score their damage level on a scale of 1 (not-damaged) to 4
(destroyed) using a multi-class classifier. We showed that while
our proposed model demonstrates decent performance, it also
works three times faster than the fastest xView2 challenge
winning solution and over 50 times faster than the slowest
first place solution, which indicates a significant improvement
from an operational perspective. We also developed a web-
based visualizer that can display the before and after imagery
along with the model’s building damage predictions on a
custom map to allow better inspection of the impacted areas by
decision-makers. This paper outlines results of a collaboration
between Microsoft AI for Good/Humanitarian Action and 510
an initiative of the Netherlands Red Cross, to help inform
field deployments using satellite imagery and AI technologies.
Our solution outperforms stakeholder’s current baseline model
significantly in terms of inference speed and segmentation