Datasets:
gremlin97
/
RemoteSensingCorpus

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
820
transformers as the backbone rather than a convolutional
neural network and relies on a lightweight decoder for the
downstream tasks.
Graph-based models have been explored in [18] for building
damage detection solutions to capture similarities between
neighboring buildings for predicting the damage. They used
the xBD dataset for cross-disaster generalization. While their
proposed approach showed some advantages in terms of
accuracy, it did not consistently outperform the Siamese
CNN model in terms of F1 score, which would be a more
appropriate metric for imbalanced datasets. Furthermore, [19]
proposed BLDNet based on a Siamese CNN combined with
a graph node classification approach to be trained in a semi-
supervised manner to reduce the number of labeled samples
needed to obtain new predictions. They benchmarked their
approach with a semi-supervised multiresolution autoencoder
and showed performance improvements. The extremely
imbalanced distributions of the building damages are
addressed in [20] by supplementing the architecture with a new
learning strategy comprising normality-imposed data-subset
generation and incremental training. However, they propose
a two-step solution approach for building localization and
damage classification. Self-supervised comparative learning
approach has been studied in [21] to address the task without
the requirement of labeled data. Their proposed approach is an
asymmetric twin network architecture evaluated on the xBD
dataset.
In this study, we propose a Siamese approach inspired by
[4], [5] where UNet architecture is used for the building
segmentation task and UNet’s encoders with shared parameters
for pre-disaster and post-disaster imagery, are used to
score building damage levels via an end-to-end approach.
Furthermore, we also evaluate the performance of our model
in various scenarios that resemble operational emergency
conditions. Web visualizer tools have been developed for
other specific domains like data-driven wildfire modeling
[22] and fire inspection prioritization [23] in the past. Our
developed web visualizer allows imagery and prediction layers
visualization for any disasters where before and after disaster
satellite images are available.
III. D ATA
In this study, we use the xBD dataset introduced in [24]
as a new large-scale dataset for the advancement of change
detection and building damage assessment for humanitarian
assistance and disaster recovery research. This dataset has been
sourced from the Maxar/DigitalGlobe Open Data Program. It
covers 19 different disasters from around the world for which
there exists high-resolution ( <0.8m/px resolution) imagery.
The disaster types include flood, wind, fire, earthquake,
tsunami, and volcano. The entire dataset contains 22,068
image tiles of 1024 ×1024 pixels that cover a total of 45,361.79
sq. km. There are 850,736 building polygons available along
with a damage level label that indicates: no-damage, minor-
damage, major-damage, and destroyed. The breakdown of the
number of polygons for pre-disaster images across different
disasters is shown in Table I. Figure 1 and Figure 2 show
some examples of pre- and post-disaster image frames from
the xBD dataset. See figure 5 for legend.
Name/Location Type # of polygons
Palu, Indonesia Earthquake/Tsunami 55,789
Mexico City, Mexico Earthquake/Tsunami 51,473
Nepal Flood 43,265
Hurricane Harvey, USA Flood 37,955
Hurricane Michael, USA Wind 35,501
Hurricane Matthew, USA Wind 23,964
Portugal Wildfire 23,413
Moore, OK Wind 22,958
Santa Rosa, CA Wildfire 21,955
SoCal, CA Wildfire 18,969
Sunda Strait, Indonesia Earthquake/Tsunami 16,847
Joplin, MO Wind 15,352
Tuscaloosa, AL Wind 15,006
Midwest USA Flood 13,896
Hurricane Florence, USA Flood 11,548
Woolsey, CA Wildfire 7,015
Pinery, Australia Wildfire 5,961
Lower Puna, HI V olcanic eruption 3,410
Guatemala V olcanic eruption 991
TABLE I
DISASTER EVENTS IN THE X BD DATASET .
IV. M ODEL ARCHITECTURE
We propose a deep learning model that conducts both
building segmentation and damage classification tasks via
a single pipeline. Our approach has some similarities to
the proposed method in [5]. However, our architecture
is less complex as we do not incorporate any attention
Fig. 1. Imagery samples from different disasters from DigitalGlobe.
(a) Pre-disaster
(b) Post-disaster
(c) Ground Truth
Fig. 2. Imagery samples with polygons showing building edges and colors
showing damage level.
layers in the model. One module of our model is based
on the UNet architecture proposed in [25], which obtains
the building segmentation mask. A single image frame is
fed to the fully convolutional UNet model where local
information is captured via encoder-decoder structures and
global information is captured via several skip connections.
In the damage assessment scenario, we have a pair of pre-
and post-disaster image frames, which are given as inputs