Semi-Supervised Domain Adaptation for Wildfire Detection
JooYoung Jang, Youngseo Cha, Jisu Kim, SooHyung Lee, Geonu Lee, Minkook Cho, Young Hwang, Nojun Kwak
May 11, 2024
Slide 1 — Frame
Wildfire Statistics
2023 USA & Hawaii Wildfire
In the USA for 2023: 34.1K fires burned 15.1M acres. The Lahaina, Hawaii wildfire (Aug 2023) caused $4-6B in damages, 115 deaths, destroyed 2K homes across 2.7K acres. Early wildfire detection could prevent such damages.
Key Statistics
34.1K Fires / 15.1M Acres
USA 2023
NIFC
Source
$4-6B Damages / 115 Deaths
Hawaii 2023
2K homes
2.7K acres
Source
National Interagency Fire Center
NCEI / NOAA
2
company@email.com
@companyusername
Slide 2 — Frame
Damage caused by wildfires
3
company@email.com
@companyusername
Lahaina, Hawaii, Aug. 8, 2023
$
2K homes
$ 4 ~ 6 billion
115 death
2.7K acres
➜ Early Wildfire Detection could prevent such damages
Slide 3 — Frame
The Problem
Why Semi-Supervised Domain Adaptation?
Domain shift occurs between training & testing environments. Limited number of target domain labeled images are available. We enhance performance by combining state-of-the-art SSL and UDA algorithms for semi-supervised domain adaptation.
4
company@email.com
@companyusername
Slide 4 — Frame
HPWREN Dataset
Open-source Data
Fixed-view cameras from HPWREN in S. California. 101 cameras with high-resolution images. 342 directories, 27,174 total images.
Proposed Labels
283 scenes labeled (30x more than previous). Source/target domain split with 0.5%, 1.0%, 3.0% protocols for SSDA evaluation.
Bbox Labels
Various scales of wildfire labeled with bounding boxes. Previous: 9 dirs / 609 images. Proposed: 283 dirs / 2,575 images.
5
company@email.com
@companyusername
Slide 5 — Frame
LADA Method
01
Pseudo Labeling
02
Translational Variance Features (CoordConv)
03
Location Aware Domain Adaptation (LADA)
6
company@email.com
@companyusername
Slide 6 — Frame
Objective Functions
Loss
Description
Supervised Loss
Use labeled source data
Masked Consistency Loss
Align masked & unmasked image output
Adversarial Loss
Align source and target domain
Consistency Loss
Align instance & image level predictions
CoordConv Layer
Replaced convolution layers in RPN and FPN with coordinate convolution layers to learn translational variance features for location-dependent wildfire patterns.
Pseudo Labeling
Generate pseudo labels for both confident foreground and background images. Reduces false positive predictions especially for 0.5% and 1.0% protocols.
7
company@email.com
@companyusername
Slide 7 — Frame
Ablation study and SSDA results showing mAP / mAP@0.5 across protocols.
Results
8
company@email.com
@companyusername
Type
Methods
0.5%
1.0%
3.0%
Source-only
SADA
6.9/21.9
9.7/28.7
17.8/48.0
Source-only
LADA
7.9/24.0
10.2/31.5
18.8/48.4
SSDA
SADA
9.7/27.3
12.3/34.9
20.4/53.0
SSDA
LADA
10.0/29.1
14.0/38.0
20.9/52.3
SSDA Result (mAP / mAP@0.5)
Slide 8 — Frame
Conclusion
SSDA Benchmark
Proposed Semi-supervised Domain Adaptation benchmark for Object Detection with wildfire dataset.
LADA Baseline
Suggested robust baseline named LADA combining pseudo labeling, CoordConv, and multiple losses.
Label Contribution
Proposed labels covering 283 scenes (30x more than previous), significantly improving detection performance.
Performance
LADA outperformed SOTA Scale Aware Domain Adaptation across all SSDA protocols in mAP.
9
company@email.com
Slide 9 — Frame
Thank You
Semi-Supervised Domain Adaptation for Wildfire Detection — ICLR 2024
