Characterizing the Spatial Distribution of Low Altitude Wind Velocity Structures in Rural Areas Following Windstorms

David Bukowski Author

07/26/2021 Added

18 Plays

Description

This projects focuses on creating an algorithm that effectively identifies tree fall patterns to help model wind directions in extreme windstorms.

Searchable Transcript

Search:

[00:00:01.488]Hello, my name is David Bukowski.
[00:00:03.968]I'm an undergraduate student at Lafayette College,
[00:00:06.195]and this is my second year working with
[00:00:07.913]Dr. Richard Wood.
[00:00:09.338]Today I'll be discussing the research
[00:00:10.596]we conducted with PhD Student Mitra Nasimi
[00:00:13.173]on post-natural event damage
[00:00:15.376]in natural areas.
[00:00:16.989]This summer, our research group analyzed
[00:00:18.941]patterns in fallen trees at post-storm sites
[00:00:21.797]to model the movement of windstorms.
[00:00:23.825]Our research focused entirely on
[00:00:25.777]how wind moves at low altitudes
[00:00:27.962]in forested areas.
[00:00:29.460]Wind moves differently depending on its
[00:00:31.093]altitude, land classification or use,
[00:00:33.710]and the topography of the land below,
[00:00:35.893]so there is the need for a wind velocity
[00:00:37.880]structure that specifically models
[00:00:40.013]near-ground, forested wind movement.
[00:00:43.221]Models for wind movement in low-altitude,
[00:00:45.565]forested areas already exist.
[00:00:47.899]However, there are multiple factors
[00:00:49.982]that make the creations of these models
[00:00:51.366]a slow and complicated process.
[00:00:53.949]Obtaining the data is time-consuming
[00:00:55.499]and expensive, as is the task of
[00:00:57.467]processing the data.
[00:00:59.199]Substantial projects survey
[00:01:00.600]tens of thousands of trees
[00:01:02.149]which then must each be processed manually.
[00:01:04.499]The outdatedness of these current methods
[00:01:06.416]is precisely why our research team decided
[00:01:08.683]to improve the treefall identification
[00:01:10.583]process.
[00:01:13.323]The main objective of this project
[00:01:15.036]is to identify fallen trees efficiently
[00:01:17.086]and objectively.
[00:01:18.569]We want to be able to process data
[00:01:20.092]as quickly and with as little human bias
[00:01:21.953]as possible.
[00:01:23.485]At the end of this research,
[00:01:24.489]we will be able to provide
[00:01:25.402]an efficient and objective method
[00:01:27.232]for identifying and quantifying
[00:01:29.045]near-ground velocity structures
[00:01:30.831]of storm winds.
[00:01:32.950]Our process begins with data collection.
[00:01:35.140]Drone use replaces traditional
[00:01:36.874]collection methods, as they can collect
[00:01:38.640]high-quality data with low cost
[00:01:40.607]in a short time.
[00:01:42.224]Our group collected the data using a
[00:01:43.657]Wingtra One drone, which returns a
[00:01:45.440]very large-scale orthomosaic map with
[00:01:47.490]high-resolution from many smaller images.
[00:01:50.144]This is done using structure-from-motion.
[00:01:53.785]We began processing the data by dividing
[00:01:56.465]each orthomosaic in 1024x1024 pixel tiles.
[00:02:00.947]We accomplished this using an algorithm
[00:02:02.880]developed in MATLAB.
[00:02:04.747]Each tile file was named according to
[00:02:06.498]a grid pattern for easy referencing.
[00:02:09.111]Once we tiled the orthomosaic,
[00:02:11.197]we needed to classify the tiles
[00:02:12.639]depending on what they depicted.
[00:02:14.744]We used a convolutional neural network
[00:02:16.610]we had developed last summer
[00:02:17.810]to sort every image into
[00:02:19.094]one of three folders:
[00:02:20.311]Treefall, Tree, or None.
[00:02:23.144]Only images classified as treefall are of
[00:02:25.328]further interest to us, but we did
[00:02:26.944]check the Tree and None folders
[00:02:28.628]for any images of treefall that were
[00:02:30.411]misclassified.
[00:02:32.798]Once we determined which images
[00:02:34.225]contained treefall, we began annotating
[00:02:35.941]these images.
[00:02:37.349]Annotation involves manually outlining
[00:02:39.438]the trunk of each fallen tree
[00:02:40.988]with a polyline.
[00:02:42.343]We use a method called
[00:02:43.137]polygon vector annotation.
[00:02:45.334]Compared to other types of annotation,
[00:02:46.936]vector annotation requires
[00:02:48.285]less precise effort
[00:02:49.334]and yields accurate results.
[00:02:51.551]Once our team annotates
[00:02:52.868]thousands of images,
[00:02:54.101]we feed them to a deep learning algorithm.
[00:02:56.001]The algorithm teaches itself
[00:02:57.385]with these inputs.
[00:02:58.868]Deep learning algorithms require
[00:03:00.301]a large amount of data
[00:03:01.468]to consistently yield good results.
[00:03:03.618]However, once our deep learning algorithm
[00:03:05.451]is given enough data,
[00:03:06.951]it will perform the same tasks that we do,
[00:03:08.818]but at a fraction of the time.
[00:03:10.990]Ultimately, this algorithm should be able
[00:03:12.885]to consistently perform
[00:03:14.301]a certain kind of classification
[00:03:15.818]called instance segmentation.
[00:03:18.202]In instance segmentation,
[00:03:20.134]the algorithm can identify
[00:03:21.752]which specific pixels of an image
[00:03:23.535]make up a fallen tree
[00:03:24.952]and the algorithm can differentiate
[00:03:26.451]between different fallen trees.
[00:03:28.372]This is incredibly important
[00:03:29.690]in post-natural event areas
[00:03:31.340]with highly-concentrated forests,
[00:03:33.007]where trees often fall on top of each other.
[00:03:36.879]This stage of our research
[00:03:37.785]has been ongoing for approximately
[00:03:39.602]a year and it continues to improve
[00:03:41.602]as we add additional annotated images.
[00:03:44.091]Here's what we expect when our
[00:03:45.369]instance segmentation algorithm
[00:03:46.720]is at our desired level.
[00:03:48.527]The algorithm will be able to analyze
[00:03:50.089]multiple square miles of terrain
[00:03:51.531]and tell us the exact number
[00:03:53.131]of fallen trees,
[00:03:54.297]the exact location of every tree,
[00:03:55.836]and the direction of each fallen tree.
[00:03:59.791]So why is this work important?
[00:04:01.701]With this information,
[00:04:02.917]we will be able to create
[00:04:04.018]a map of post-natural event areas
[00:04:05.517]where each fallen tree is represented
[00:04:07.184]as a vector.
[00:04:08.524]While these vector maps already do exist
[00:04:10.291]for a very few number of events,
[00:04:11.875]they are made entirely by hand
[00:04:13.508]and take a large amount
[00:04:15.008]of humanpower and time as a result.
[00:04:17.591]A map generated with a
[00:04:18.858]deep learning algorithm
[00:04:20.041]would allow for researchers
[00:04:21.061]to analyze near-ground storm
[00:04:22.775]wind velocities more efficiently
[00:04:24.225]and objectively than ever before.
[00:04:27.225]Thank you for taking the time
[00:04:28.975]to listen to my presentation.
[00:04:30.810]I would like to thank Dr. Wittich,
[00:04:32.125]and Dr. Bartelt-Hunt for
[00:04:33.259]running a great program this summer.
[00:04:35.341]Thank you also to UNL's
[00:04:36.508]Civil & Environmental Engineering
[00:04:38.042]department and UNL's graduate
[00:04:39.326]studies program for making sure
[00:04:40.875]students got the most
[00:04:42.092]out of this summer.
[00:04:43.775]Thank you to the National Science Foundation,
[00:04:45.192]The Northern Tornados project,
[00:04:46.709]and the Structural Extreme Events
[00:04:48.475]Reconnaissance Network for
[00:04:49.708]funding this project and providing data.
[00:04:52.358]Finally, I would especially like to thank
[00:04:54.059]Dr. Richard Wood and Ms. Mitra Nasimi
[00:04:56.314]for all of their help and input
[00:04:57.492]this summer.
[00:04:58.625]It's been an absolute pleasure
[00:04:59.563]working with them.
[00:05:00.368]Thank you for listening.

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

Characterizing the Spatial Distribution of Low Altitude Wind Velocity Structures in Rural Areas Following Windstorms

Description

Searchable Transcript

Comments icon comment

Related Channels

Comments