Multiclass Semantic Segmentation of Wildland Fire Images Using Context-Aware Centralized Copy-Paste Data Augmentation
Authors: Joon Tai Kim, Nishanth Kunchala, Vishv Patel, Tianle Chen, Ziyu Dong, Daniel Ospina Acero, Roger Williams, Mrinal Kumar
Organizations: Mechanical and Aerospace Engineering Department, The Ohio State University, Columbus, OH 43210, USA · Department of Computer Science, Boston University, Boston, MA 02215, USA · Department of Wildland Resources, Utah State University, Logan, UT 84322, USA · Department of Electronic and Telecommunications Engineering, Universidad de Antioquia, Medell´ın, Colombia · School of Environment and Natural Resources, The Ohio State University, Columbus, OH 43210, USA
Producing accurate annotations for deep learning based image segmentation is both costly and labor intensive. This challenge is especially evident in wildland fire applications, where accurately labeled datasets are scarce due to the difficulty of collecting and annotating dynamic fire scenes. To address this problem, our previous work introduced the Centralized Copy-Paste Data Augmentation (CCPDA) method for semantic segmentation of wildland fire imagery, which generates artificial training samples by randomly pasting fire clusters from source images onto target images. However, random placement can produce contextually unrealistic scenes, such as fire burning on asphalt. In this paper, we present a context-aware strategy designed specifically to improve data quality and realism in small multiclass wildland fire datasets, ensuring that augmented samples remain contextually meaningful. The proposed method restricts fire placement to semantically valid target regions and selects the location whose Ash-Vegetation composition most closely matches the source context. This approach preserves existing fire regions in the target image, prevents unrealistic placements, and maintains contextual accuracy by generating images that resemble real wildland fire scenes. We evaluate the Context-Aware CCPDA strategy through numerical analysis and comparisons with other augmentation methods by a weighted sum-based multi-objective optimization (MOO) approach. The results confirm that the context-aware data augmentation strategy leads to improved segmentation performance and contextual realism, outperforming other augmentation procedures.
Figures & tables
Fig. 1 : The U-Net architecture used as the baseline model for semantic segmentation of wildland fire images. This figure is reused from [ 10 ] .
Fig. 2 : Flowchart for smoke dehazing method. This figure is reused from [ 10 ] .
Fig. 3 : A comparison of the original hazy RGB image (a) and the dehazed RGB image (b) achieved through a 95% haze-reduction procedure.
Fig. 4 : An illustration of the standard centralized copy-paste method for fire segmentation. The process involves: (1) identifying an original fire cluster (yellow) in the source image, (2) applying the centralized method to focus on the core fire area (red), and (3) pasting the centralized fire cluster onto a target image. This figure is reused from [ 10 ] .
Fig. 5 : Flowchart illustrating the proposed Context-Aware CCPDA method. The process identifies fire clusters, computes Ash and Vegetation statistics, and evaluates candidate paste locations in the target image based on compositional similarity and semantic constraints. Finally, the optimal location is selected; if none is identified, the bounding box is adaptively downsized and the process repeated.
Fig. 6 : A visual comparison of Non-Context-Aware and Context-Aware CCPDA results for a sample source and target image pair. Sub-figures (a) and (b) show the original source and target images, respectively, while (c) and (d) display zoomed-in augmented images from the upper-center region of (b), generated using the Non-Context-Aware and Context-Aware CCPDA methods, respectively.
Fig. 7 : A visualization of compositional matching for wildland fire image data augmentation. Green and yellow data points represent Ash –to- Vegetation percentage distributions from the source and target images, respectively. The inset illustrates the context-aware paste-location selection process for a single source fire cluster, where candidate target bounding boxes (red points) are compared with the source bounding box (single blue point), and the target bounding box with the closest compositional match is selected (single green star).
Learning Rate
Dropout Rate
Batch Size
F(x)
0.0005
0.3
4
0.83734
0.0005
0.2
16
0.83650
0.0005
0.2
4
0.83570
⋮
0.005
0.3
8
0.23578
0.01
0.1
4
0.23578
TABLE I : Selected hyperparameter configurations ranked by F(x) score. The first row, shown in bold, indicates the best-performing tuned configuration.
Method
Fire -FNR
Veg. -IoU
total-IoU
F(x)
Non-Augmented
18.04
67.40
54.45
0.7486
Rotation-by-18 ∘
14.17
71.14
60.54
0.7894
Non-Context-Aware Copy-Paste
8.24
71.52
60.91
0.8271
Context-Aware Copy-Paste
6.75
73.74
64.03
0.8458
TABLE II : Study 1 performance metrics (%) and weighted-sum scores across different augmentation methods.
(Study 1) Method
Fire [%]
(Study 2) Erosion %
Fire [%]
Non-Augmented
5.07
0%
8.80
Rotation-by-18 ∘
10.59
10%
8.31
N.C.A 1 Copy-Paste
9.30
20%
7.79
C.A 2 Copy-Paste
9.19
30%
7.38
Non-Augmented Data Pixel Total
655,360
Augmented Data Pixel Total
13,107,200
TABLE III : Comparison of Fire pixel percentages across Study 1 augmentation methods and Study 2 erosion levels for the expanded 26-image set, including total pixel counts for the corresponding non-augmented and augmented training datasets. 1 Non-Context-Aware, 2 Context-Aware.
Fig. 8 : A visual comparison of the semantic segmentation results of Study 1 for a sample test RGB image.
Fig. 9 : A visual comparison of Context-Aware CCPDA outputs using the first and a randomly selected optimal location for a sample training image. Sub-figure (b) shows the training image used as both the source and target image. Sub-figures (d) and (e) show the Context-Aware CCPDA outputs obtained by pasting the source fire clusters at the first optimal location and at a randomly selected location from the set of possible optimal locations in the target image, respectively. Sub-figures (a) and (c) provide enlarged views of the corresponding boxed regions in (d) and (e), respectively.
20-image set
26-image set
Method
Fire -FNR
Veg. -IoU
total-IoU
F(x)
Fire -FNR
Veg. -IoU
total-IoU
F(x)
Erosion-by-0%
5.22
66.98
56.16
0.8276
6.76
70.32
60.45
0.8323
Erosion-by-10%
4.00
66.79
57.93
0.8365
4.67
70.78
60.17
0.8460
Erosion-by-20%
7.59
67.99
57.29
0.8172
8.28
70.34
58.88
0.8213
Erosion-by-30%
9.04
68.15
57.09
0.8086
12.27
72.31
62.05
0.8059
TABLE IV : Study 2 performance metrics (%) and weighted-sum scores across erosion levels for the Context-Aware CCPDA method on the original 20-image set used to construct BURN 1 and the expanded 26-image set used to construct BURN 2.
Method
Fire -FNR
Veg. -IoU
total-IoU
F(x)
FCN
Non-Augmented
72.85
65.23
49.72
0.4024
Rotation-by-18 ∘
25.35
66.82
54.67
0.7026
Non-Context-Aware Copy-Paste
22.54
67.33
54.34
0.7208
Context-Aware Copy-Paste
21.72
66.19
52.99
0.7211
SegNet
TABLE V : Study 3 performance metrics (%) and weighted-sum scores across different augmentation methods on three semantic segmentation models: FCN, SegNet, and DeepLabV3+.
The scarcity of labeled satellite imagery remains a fundamental bottleneck for deep-learning (DL)-based wildfire monitoring systems. This paper investigates whether a diffusion-based foundation model for Earth Observation (EO), EarthSynth, can synthesize realistic post-wildfire Sentinel-2 RGB imagery conditioned on existing burn masks, without task-specific retraining. Using burn masks derived from the CalFireSeg-50 dataset (Martin et al., 2025), we design and evaluate six controlled experimental configurations that systematically vary: (i) pipeline architecture (mask-only full generation vs. inpainting with pre-fire context), (ii) prompt engineering strategy (three hand-crafted prompts and a VLM-generated prompt via Qwen2-VL), and (iii) a region-wise color-matching post-processing step. Quantitative assessment on 10 stratified test samples uses four complementary metrics: Burn IoU, burn-region color distance (ΔC_burn), Darkness Contrast, and Spectral Plausibility. Results show that inpainting-based pipelines consistently outperform full-tile generation across all metrics, with the structured inpainting prompt achieving the best spatial alignment (Burn IoU = 0.456) and burn saliency (Darkness Contrast = 20.44), while color matching produces the lowest color distance (ΔC_burn = 63.22) at the cost of reduced burn saliency. VLM-assisted inpainting is competitive with hand-crafted prompts. These findings provide a foundation for incorporating generative data augmentation into wildfire detection pipelines. Code and experiments are available at: https://www.kaggle.com/code/valeriamartinh/genai-all-runned
Valeria Martin, K. Brent Venable, Derek Morgan
Department of Intelligent Systems and Robotics, University of West Florida, Pensacola, FL, USA · IHMC, Pensacola, FL, USA
Predicting future states in uncertain environments, such as wildfire spread, medical diagnosis, or autonomous driving, requires models that can consider multiple plausible outcomes. While diffusion models can effectively learn such multi-modal distributions, naively sampling from these models is computationally inefficient, potentially requiring hundreds of samples to find low-probability modes that may still be operationally relevant. In this work, we address the challenge of sample-efficient ambiguous segmentation by evaluating several training-free sampling methods that encourage diverse predictions. We adapt two techniques, particle guidance and SPELL, originally designed for the generation of diverse natural images, to discrete segmentation tasks, and additionally propose a simple clustering-based technique. We validate these approaches on the LIDC medical dataset, a modified version of the Cityscapes dataset, and MMFire, a new simulation-based wildfire spread dataset introduced in this paper. Compared to naive sampling, these approaches increase the HM IoU* metric by up to 7.5% on MMFire and 16.4% on Cityscapes, demonstrating that training-free methods can be used to efficiently increase the sample diversity of segmentation diffusion models with little cost to image quality and runtime. Code and dataset: https://github.com/SebastianGer/wildfire-spread-scenarios
Sebastian Gerard, Josephine Sullivan
KTH Royal Institute of Technology, Stockholm, Sweden
This article describes an open image dataset for developing and evaluating active-fire segmentation methods in satellite imagery. The dataset contains 2,148 image-mask pairs from 25 California wildfires, with acquisitions spanning July 2020 to August 2026. Each image is a 512x512-pixel, three-channel composite derived from Sentinel-2 Level-2A bands B12, B11 and B8A at 20 m spatial sampling. A fixed linear rendering is applied throughout the dataset. Corresponding masks distinguish background, SWIR-rule active fire and invalid observations. The masks were generated from shortwave-infrared brightness and near-infrared contrast, followed by constrained neighborhood growth. The release includes chip-level metadata and an incident-disjoint partition containing 18 training, three validation and four test fires. Among the image pairs, 841 contain active-fire labels; these labels occupy 0.0766% of all grid cells. A mask-blind analyst review covers 233 test chips and provides a separate assessment of the rule-generated labels at chip and connected-component levels. Reference training and evaluation code accompanies the data, including a ResNet-34 U-Net implementation with validation-based checkpoint and threshold selection. The archived images, masks, metadata and review annotations support research on rare-class segmentation, learning from algorithmic labels and transfer across fire incidents. The versioned dataset is deposited on Zenodo, with preparation and reuse software maintained in a public GitHub repository.
California High School, San Ramon, CA, 94583, USA · Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, 95616, USA