Assessment of Post-Tornado Damage in Natural Areas

David Bukowski Author

07/26/2020 Added

19 Plays

Description

A 5-minute overview of David Bukowski's summer research.

Searchable Transcript

Search:

[00:00:00.000]Hello, my name is David Bukowski.
[00:00:02.355]I’m an undergraduate student at Lafayette College.
[00:00:04.995]Today I’ll be discussing my research
[00:00:07.167]on post-tornado damage in natural areas
[00:00:09.492]that I conducted with Dr. Richard Wood and Yijun Liao.
[00:00:14.922]This summer, my research group and I
[00:00:16.939]analyzed patterns in fallen trees
[00:00:18.969]at post-tornado sites
[00:00:20.319]in order to model how tornado winds act.
[00:00:23.599]My research focuses entirely
[00:00:25.050]on how these winds move
[00:00:26.234]at low altitudes in forested areas.
[00:00:28.584]Wind moves differently
[00:00:29.669]depending on its altitude
[00:00:30.820]and on the topography of the land below,
[00:00:33.209]so there is the need for
[00:00:34.492]a wind velocity structure
[00:00:35.695]that specifically models near-ground,
[00:00:37.476]forested wind movement.
[00:00:39.958]Models for wind movement
[00:00:41.106]in low-altitude, forested areas
[00:00:42.615]already exist.
[00:00:44.020]However, there are multiple factors
[00:00:45.758]that make the creation of these models
[00:00:47.508]a slow and manually-intensive process.
[00:00:50.825]Obtaining the data is
[00:00:52.061]time-consuming and expensive,
[00:00:53.545]as is the task of processing the data.
[00:00:56.360]Substantial projects survey tens of thousands of trees
[00:00:59.215]which then must each be processed by hand.
[00:01:02.906]The outdatedness of these current methods
[00:01:04.769]are exactly why our research team
[00:01:06.414]decided to improve the
[00:01:07.554]treefall identification process.
[00:01:10.717]The main objective of this project
[00:01:12.370]is to identify fallen trees efficiently and objectively.
[00:01:15.498]We want to be able to process data
[00:01:17.534]as quickly and with as little human bias as possible.
[00:01:21.045]Should we succeed in doing this,
[00:01:22.932]we will be able provide an efficient and objective method
[00:01:25.790]for modeling near-ground velocity structures of tornado winds.
[00:01:30.016]To reduce the time needed for data collection,
[00:01:33.033]we used Structure-from-Motion technology
[00:01:35.263]attached to a drone
[00:01:36.633]to obtain our data.
[00:01:38.049]This method produces cheaper and quicker results,
[00:01:41.080]capturing a hectacre of land every minute
[00:01:43.668]or multiple square miles every day.
[00:01:46.234]From this data, we can attain both
[00:01:48.367]a 3 dimensional point cloud model of the disaster area
[00:01:51.368]and a 2 dimensional model derived from the point cloud.
[00:01:56.546]Deep learning is the main focus in our effort
[00:01:58.936]to make the wind modeling process
[00:02:00.997]more effective, quick, and objective.
[00:02:04.167]Deep learning is a subset of machine learning
[00:02:06.827]that learns automatically when fed raw data.
[00:02:10.097]For the final step of this project,
[00:02:11.777]we want to use a complex,
[00:02:12.756]intensive method of deep learning
[00:02:14.656]called instance segmentation.
[00:02:16.526]However, to understand
[00:02:18.063]how to best optimize the final algorithm,
[00:02:20.549]we first experimented
[00:02:21.988]with a much simpler type of deep learning
[00:02:23.901]called classification.
[00:02:25.937]To save time, we took an already-existing
[00:02:28.760]deep learning classification algorithm
[00:02:30.777]and taught it to learn how to identify fallen trees.
[00:02:33.802]We manually labelled 500 aerial images
[00:02:36.760]from our post-tornado survey
[00:02:38.801]as either a tree, a fallen tree, or no tree at all,
[00:02:41.977]and then fed each labelled image to the algorithm.
[00:02:45.527]The algorithm then learned to classify an image
[00:02:47.644]as a tree, a fallen tree, or no tree at all.
[00:02:51.060]To understand how to optimize an algorithm,
[00:02:53.928]we created 6 different augmentation sets
[00:02:57.551]The first set contained no augmentation,
[00:02:59.477]while the other 5 consisted of combinations
[00:03:02.011]of 4 augmentation functions,
[00:03:03.678]reflection, rotation, shear, and translation.
[00:03:07.578]We knew each of these functions affected the performance
[00:03:10.744]of deep learning algorithms in some manner,
[00:03:13.144]but we needed to run tests to determine
[00:03:15.745]which combination of functions was the most effective.
[00:03:19.511]We ran our classification algorithm
[00:03:21.303]multiple times with each augmentation set,
[00:03:23.377]using the same 500 images each time.
[00:03:26.728]We analyzed how well the algorithm learned
[00:03:28.677]with each augmentation set
[00:03:30.394]and recorded the validation accuracy
[00:03:32.711]and validation loss from every trial.
[00:03:35.194]The results showed that the second set of augmentations,
[00:03:37.444]a combination of rotation and reflection functions,
[00:03:40.228]made the algorithm run most effectively.
[00:03:44.302]Since we identified the most effective parameters
[00:03:46.958]for our deep learning algorithms,
[00:03:48.487]we moved on to work on the more
[00:03:50.422]complex type of deep learning,
[00:03:52.029]instance segmentation.
[00:03:54.488]In instance segmentation,
[00:03:55.854]the algorithm can identify
[00:03:57.454]which specific pixels of an image
[00:03:59.288]make up a fallen tree,
[00:04:00.904]and the algorithm can differentiate
[00:04:02.555]between different fallen trees.
[00:04:04.721]Like our first algorithm,
[00:04:06.092]we used transfer learning
[00:04:07.277]for this algorithm as well,
[00:04:08.705]teaching an already existing algorithm
[00:04:10.671]to identify fallen trees.
[00:04:12.854]Unfortunately, this stage of our research is not finished yet.
[00:04:16.138]However, when our instance segmentation algorithm is complete,
[00:04:19.306]here’s what we should expect.
[00:04:21.855]The algorithm will be able to analyze
[00:04:23.273]multiple square miles of terrain
[00:04:24.971]and tell us the exact number of fallen trees,
[00:04:27.521]the exact location of every tree,
[00:04:29.605]the direction of each fallen tree,
[00:04:32.722]and provide a bounding box for each tree.
[00:04:37.139]So why is this important?
[00:04:39.012]Well, with this information,
[00:04:40.329]we will be able to create a map
[00:04:41.662]of post-disaster areas
[00:04:43.112]where each fallen tree
[00:04:44.716]is represented as a vector.
[00:04:46.757]Such a map would allow for researchers
[00:04:48.740]to analyze near-ground tornado wind velocities
[00:04:51.006]more efficiently and objectively
[00:04:52.883]than ever before.
[00:04:55.341]Thank you for taking the time to listen to my presentation.
[00:04:57.738]I would like to thank Dr. Wittich
[00:05:00.061]for organizing the sustainability REU this summer
[00:05:02.811]and adapting to all the changes that she had to deal with.
[00:05:05.827]Thank you also to UNL’s Civil Engineering department
[00:05:08.577]and UNL’s graduate studies program
[00:05:10.912]for making sure students got the most out of this summer.
[00:05:13.911]Thank you to the National Science Foundation
[00:05:15.594]and StEER for funding this project and providing data.
[00:05:18.879]Finally, I would especially like to thank
[00:05:21.611]Dr. Richard Wood and PhD candidate Yijun Liao
[00:05:24.876]for all of their help and input this semester.
[00:05:27.000]this work would not have been possible without them.
[00:05:29.733]Thank you

The screen size you are trying to search captions on is too small!

You can always jump over to MediaHub and check it out there.

Comments

0 Comments

Assessment of Post-Tornado Damage in Natural Areas

Description

Searchable Transcript

Comments icon comment

Related Channels

Comments