-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathLandslide.html
130 lines (115 loc) · 6.39 KB
/
Landslide.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
<!DOCTYPE HTML>
<!--
Hyperspace by HTML5 UP
html5up.net | @ajlkn
Free for personal and commercial use under the CCA 3.0 license (html5up.net/license)
-->
<html>
<head>
<title>MUTATED</title>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1, user-scalable=no" />
<link rel="stylesheet" href="assets/css/main.css" />
<noscript><link rel="stylesheet" href="assets/css/noscript.css" /></noscript>
</head>
<body class="is-preload">
<!-- Header -->
<header id="header">
<a href="index.html" class="title">Jenna Abrahamson</a>
<nav>
<ul>
<li><a href="index.html">Home</a></li>
</ul>
</nav>
</header>
<!-- Wrapper -->
<div id="wrapper">
<!-- Main -->
<section id="main" class="wrapper">
<div class="inner">
<h1 class="major">MUTATED</h1>
<p>
<h2> Modeling and Understanding using Temporal Analysis of Transient Earth Data </h2>
<span class="image fit"><img src="images/SMART.png" alt="" /></span>
As a part of my research assistantship, I am a member of the MUTATED team. The MUTATED team is part of
IARPA's (Intelligence Advanced Research Projects Activity) Space-Based Machine Automated Recognition Technique
(<a href="https://www.iarpa.gov/research-programs/smart" class="title">SMART</a>) Program working to automate broad-area
search of multi-source satellite imagery to detect, monitor, and characterize the progression of anthropogenic or
natural processes.
<br>
<br>
To address this goal, our team has been working extensively on the development of a novel online change detection algorithm.
This algorithm, termed <strong>roboBayes</strong>, was based on the established Bayesian Online Changepoint
Detection Algorithm (BOCPD, <a href="https://arxiv.org/abs/0710.3742" class="title">Adams and MacKay 2007</a>) but created
and implemented specifically for remote sensing change detection by team member Laura Wendelberger
(<a href="https://arxiv.org/abs/2112.12899?context=stat" class="title">Wendelberger et. al 2021</a>). My work on this
algorithm is outlined below:
<br>
</p>
<h2> Outlier Detection </h2>
<p> To help address the issue of outliers in our data, due to clouds or other sources of atmospheric interference, I implemented
two different spike filter options in the pre-processing portion of the roboBayes pipeline. The first, is a lagging spike filter
which uses a moving window of the previous values compared to the next data observation in order to flag a spike, or outlier,
that is outside of a set standard deviation threshold. The second spike filter, is a traditional moving window spike filter
where the moving window looks at the center data point of the window and flags a spike, or outlier, that is outside of a set standard
deviation threshold.
</p>
<span class="image fit"><img src="images/Outlier.png" alt="" /></span>
<br>
<h2> Heuristics-Based Filtering </h2>
<p> To aid in eliminating false positives and classifying changes related to heavy construction, I developed a heuristics-based
filter in the post-processing portion of the roboBayes pipeline. To determine what remote sensing signals and model
coefficients were indicative of heavy construction, I performed unsupervised K-means clustering of all
model coefficients across multiple regions for all signals that contained a change point. These coefficients included the model mean,
amplitude, and variance for each signal. This helped me to target signals that were most important for heavy construction
classification to create a parameter filter that improved our precision across regions while still maintaining high levels
of recall.
<br>
</p>
<figure>
<span class="image fit"><img src="images/Filter.png" alt="" /></span>
<figcaption><i>This image shows for two construction change sites (annotated in red) how patterns emerge in the spectral
model values for identified change pixels. This information was used to build parameter filters targeted at
classifying construction from other types of change.</i></figcaption>
</figure>
<br>
<br>
<h2> Tuning: Paramter Grid Search </h2>
<p> To tune the roboBayes algorithm and find the best parameter set, I performed a grid search and set of tuning runs
for the spike filter and between the spike filter and observation variance. This was performed across multiple
regions and geographic terrains to try and find the best generalizable parameter set. Runs were completed using
the NCSU High-Performance Computing (HPC) cluster.
<br>
</p>
<figure>
<span class="image fit"><img src="images/Tune.gif" alt="" /></span>
<figcaption><center><i>This animation shows a set of 12 tuning runs performed for change detection
results for the island of Bahrain.</i></center></figcaption>
</figure>
<br>
<h3>Associated Articles:</h3>
<ul>
<li><a href="https://cnr.ncsu.edu/geospatial/news/2021/02/08/new-research-will-improve-accuracy-efficiency-of-imagery-analysis/">New IARPA-Funded Research Will Improve Accuracy and Efficiency of Analyzing Geospatial Imagery Across Large Scales</a></li>
<li><a href="https://www.wired.com/story/spotting-objects-from-space-is-easy-this-challenge-is-harder/">Spotting Objects from Space is Easy. This Challenge is Harder. </a></li>
</ul>
</div>
</section>
</div>
<!-- Footer -->
<footer id="footer" class="wrapper alt">
<div class="inner">
<ul class="menu">
<li>© Untitled. All rights reserved.</li><li>Design: <a href="http://html5up.net">HTML5 UP</a></li>
</ul>
</div>
</footer>
<!-- Scripts -->
<script src="assets/js/jquery.min.js"></script>
<script src="assets/js/jquery.scrollex.min.js"></script>
<script src="assets/js/jquery.scrolly.min.js"></script>
<script src="assets/js/browser.min.js"></script>
<script src="assets/js/breakpoints.min.js"></script>
<script src="assets/js/util.js"></script>
<script src="assets/js/main.js"></script>
</body>
</html>