國立虎尾科技大學 |

Building Damage Assessment Using Remote Sensing Data and Deep Learning Algorithms.

Record Type:	Language materials, manuscript : Monograph/item
Title/Author:	Building Damage Assessment Using Remote Sensing Data and Deep Learning Algorithms./
Author:	Sodeinde, Olalekan R.
Description:	1 online resource (77 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 85-08, Section: A.
Contained By:	Dissertations Abstracts International85-08A.
Subject:	Remote sensing. -
Online resource:	click for full text (PQDT)
ISBN:	9798381701623

Building Damage Assessment Using Remote Sensing Data and Deep Learning Algorithms.
Sodeinde, Olalekan R.

Building Damage Assessment Using Remote Sensing Data and Deep Learning Algorithms. - 1 online resource (77 pages)

Source: Dissertations Abstracts International, Volume: 85-08, Section: A.

Thesis (Ph.D.)--Tufts University, 2024.

Includes bibliographical references

Climate change and wars bring catastrophic disaster events, which raise the need for quicker turnaround in building damage assessments. Disaster response personnels and non-governmental and governmental agencies require information on building damage to make timely, life-saving decisions such as resource allocations. To help automate this process, the xBD dataset was released for training deep learning algorithms. The dataset includes high-resolution satellite imagery annotated with building locations and damage classes ("destroyed", "major damage", "minor damage", and "no damage") before and after natural disasters. This thesis analyzed the quality of the xBD dataset for identifying building damage across a variety of natural hazards using deep learning algorithms.In Chapter 1, models were created to evaluate the pros and cons of combining the training datasets across multiple natural hazards. Because the model overfits to the "no damage" class, merging classes was evaluated. The "no damage" class was retained and a second class ("damaged") was created by merging "minor damage," "major damage," and "destroyed" classes. Recommendations were then made on using the provided training dataset for optimizing classification accuracy for building damage assessment across hazards including: volcanoes, hurricanes, wildfires, floods, tsunamis, earthquakes, tornadoes, and fires.Chapter 2 operationalizes the models from Chapter 1 for a new, unseen natural disaster event. It followed recommendations from Chapter 1 to select binary and multiclass models to perform damage assessments of the town of 'Eua, Tonga. The buildings in 'Eua, Tonga were affected by the January 2022 volcanic eruption and its resulting tsunami. However, the damages to the buildings were primarily caused by the tsunami. Post-disaster high-resolution (resolution of 0.8 m) satellite images from the Worldview-3 satellite (provided by Maxar Technologies through their the National Geospatial-Intelligence Agency (NGIA) access) were used with a pre-trained deep learning classification algorithm to classify the damages in 'Eua. Building footprint images are required for the damage classification models. The building footprints were downloaded from Open Street Maps (OSM). Validation of the results was based on United Nations Satellite Center (UNOSAT) visual interpretation of Pleiades images. We tested both a binary class and multiclass model trained using the multi-hazard xbd dataset. The result showed that the binary model trained on a multi-hazard xBD overfitted to the no-damage class. However, the multiclass model that was trained on a multi-hazard xBD dataset was transferable. It was able to detect 96% of no-damage buildings and 84% of damaged buildings.Chapter 3 operationalizes the models from Chapter 1 for a non-natural disaster event. It examined using pre-trained classification models developed in Chapter 1 for classifying post-war images of buildings damaged by Russian invasion of Bucha, Ukraine in March 2022. We evaluated models trained using wind hazards and multi-hazards on a multi-class damage data. Building footprint images were also required for the damage classification models. For completeness, missing building footprints from OSM were added. The results from the models were validated against the UNOSAT post-disaster dataset, which only contained information about damaged buildings. The results indicate that the classification model trained with wind data detected minor damage (54%), but was unable to detect the major and destroyed classes. The multi-hazard model was able to detect some of the major (18%) and minor (27%) damages, but was unable to detect any of the destroyed buildings. However, both models overfit to the "no damage" class.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798381701623Subjects--Topical Terms:

557272
Remote sensing.
Subjects--Index Terms:

Damage assessmentIndex Terms--Genre/Form:

554714
Electronic books.

Building Damage Assessment Using Remote Sensing Data and Deep Learning Algorithms.
LDR:05164ntm a22004097 4500 001 1142901
005 20240429063900.5
006 m o d
007 cr mn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798381701623
035 $a (MiAaPQ)AAI30989781
035 $a AAI30989781
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Sodeinde, Olalekan R. $3 1467423
245 1 0 $a Building Damage Assessment Using Remote Sensing Data and Deep Learning Algorithms.
264 0 $c 2024
300 $a 1 online resource (77 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-08, Section: A.
500 $a Advisor: Baise, Laurie.
502 $a Thesis (Ph.D.)--Tufts University, 2024.
504 $a Includes bibliographical references
520 $a Climate change and wars bring catastrophic disaster events, which raise the need for quicker turnaround in building damage assessments. Disaster response personnels and non-governmental and governmental agencies require information on building damage to make timely, life-saving decisions such as resource allocations. To help automate this process, the xBD dataset was released for training deep learning algorithms. The dataset includes high-resolution satellite imagery annotated with building locations and damage classes ("destroyed", "major damage", "minor damage", and "no damage") before and after natural disasters. This thesis analyzed the quality of the xBD dataset for identifying building damage across a variety of natural hazards using deep learning algorithms.In Chapter 1, models were created to evaluate the pros and cons of combining the training datasets across multiple natural hazards. Because the model overfits to the "no damage" class, merging classes was evaluated. The "no damage" class was retained and a second class ("damaged") was created by merging "minor damage," "major damage," and "destroyed" classes. Recommendations were then made on using the provided training dataset for optimizing classification accuracy for building damage assessment across hazards including: volcanoes, hurricanes, wildfires, floods, tsunamis, earthquakes, tornadoes, and fires.Chapter 2 operationalizes the models from Chapter 1 for a new, unseen natural disaster event. It followed recommendations from Chapter 1 to select binary and multiclass models to perform damage assessments of the town of 'Eua, Tonga. The buildings in 'Eua, Tonga were affected by the January 2022 volcanic eruption and its resulting tsunami. However, the damages to the buildings were primarily caused by the tsunami. Post-disaster high-resolution (resolution of 0.8 m) satellite images from the Worldview-3 satellite (provided by Maxar Technologies through their the National Geospatial-Intelligence Agency (NGIA) access) were used with a pre-trained deep learning classification algorithm to classify the damages in 'Eua. Building footprint images are required for the damage classification models. The building footprints were downloaded from Open Street Maps (OSM). Validation of the results was based on United Nations Satellite Center (UNOSAT) visual interpretation of Pleiades images. We tested both a binary class and multiclass model trained using the multi-hazard xbd dataset. The result showed that the binary model trained on a multi-hazard xBD overfitted to the no-damage class. However, the multiclass model that was trained on a multi-hazard xBD dataset was transferable. It was able to detect 96% of no-damage buildings and 84% of damaged buildings.Chapter 3 operationalizes the models from Chapter 1 for a non-natural disaster event. It examined using pre-trained classification models developed in Chapter 1 for classifying post-war images of buildings damaged by Russian invasion of Bucha, Ukraine in March 2022. We evaluated models trained using wind hazards and multi-hazards on a multi-class damage data. Building footprint images were also required for the damage classification models. For completeness, missing building footprints from OSM were added. The results from the models were validated against the UNOSAT post-disaster dataset, which only contained information about damaged buildings. The results indicate that the classification model trained with wind data detected minor damage (54%), but was unable to detect the major and destroyed classes. The multi-hazard model was able to detect some of the major (18%) and minor (27%) damages, but was unable to detect any of the destroyed buildings. However, both models overfit to the "no damage" class.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Remote sensing. $3 557272
650 4 $a Information science. $3 561178
653 $a Damage assessment
653 $a Deep learning
653 $a Disaster response
653 $a Machine learning
653 $a Neural networks
655 7 $a Electronic books. $2 local $3 554714
690 $a 0543
690 $a 0723
690 $a 0800
690 $a 0799
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a Tufts University. $b Civil Engineering. $3 1179655
773 0 $t Dissertations Abstracts International $g 85-08A.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30989781 $z click for full text (PQDT)