CCPDA: Centralized Copy-Paste Augmentation

• The paper introduces CCPDA, a method that isolates and centralizes fire cores before pasting them onto target images to improve fire segmentation.
• It reports significant improvements in fire class metrics such as IoU, precision, and recall using a weighted multi-objective evaluation framework.
• Limitations include potential physical inconsistencies and lighting issues with pasted cores, with future work proposed in context-aware placements and adaptive blending.

Centralized Copy-Paste Data Augmentation (CCPDA) is a data-augmentation method introduced for training deep-learning multiclass segmentation models in wildland fire science, with special focus on improving segmentation outcomes for the fire-class. It was presented in "Centralized Copy-Paste: Enhanced Data Augmentation Strategy for Wildland Fire Semantic Segmentation" (Kim et al., 8 Jul 2025), published on 2025-07-08. The method has three main steps: (i) identify fire clusters in the source image, (ii) apply a centralization technique to focus on the core of the fire area, and (iii) paste the refined fire clusters onto a target image. In the application scenario considered, the method increases dataset diversity while preserving the essential characteristics of the fire class, and numerical performance assessment validates its efficacy in alleviating the difficulties associated with small, manually labeled training datasets (Kim et al., 8 Jul 2025).

1. Problem setting and motivation

Wildland-fire semantic-segmentation models are described as suffering from two intertwined issues: extreme class imbalance and small, costly datasets (Kim et al., 8 Jul 2025). In the formulation given for the method, “fire” regions occupy only a few percent of each high-resolution frame, yet accurate fire delineation is operationally critical. At the same time, manually labeling fire every pixel is expensive, so publicly available training sets are tiny.

The stated consequence is that deep networks overfit background and fuel classes, under-segment the fire class, and exhibit poor generalization. CCPDA was invented to synthesize new “fire” examples by isolating authentic fire clusters (“cores”), recentralizing them, and pasting them onto diverse background scenes at training time. The reported rationale is twofold: the augmentation both balances the fire-class frequency and preserves the high-frequency textural details that distinguish fire from other classes.

The paper explicitly frames the operational objective around the fire class rather than all classes symmetrically. Fuel, ash, and background are part of the multiclass setting, but the fire class is treated as carrying significantly more operational significance than the other classes. This emphasis shapes both the augmentation design and the evaluation criterion.

2. Formal dataset model and fire-cluster extraction

The original dataset is defined as

$\mathcal{D} = \{(I_i, M_i)\}_{i=1}^n,$

where $I_i \in \mathbb{R}^{H\times W\times 3}$ are RGB images and $M_i \in \{0,\dots,C-1\}^{H\times W}$ are pixel-wise masks, with one of the $C$ labels being “fire” (Kim et al., 8 Jul 2025).

Fire Cluster Identification begins by extracting the binary fire mask

$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$

Connected components are then computed with 8-connectivity on $F_i$ to obtain segments $\{s_{i,j}\}$, formalized as

$$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$

The procedure optionally dilates each segment to include a thin rim of hot-air pixels:

$s'_{i,j} = s_{i,j} \oplus K_5,$

where $K_5$ is a $I_i \in \mathbb{R}^{H\times W\times 3}$0 square structuring element and $I_i \in \mathbb{R}^{H\times W\times 3}$1 denotes morphological dilation. A subsequent area filter retains only

$I_i \in \mathbb{R}^{H\times W\times 3}$2

with $I_i \in \mathbb{R}^{H\times W\times 3}$3 pixels recommended.

This extraction stage serves a selective role rather than a purely geometric one. By discarding tiny noise clusters and preserving a thin rim around flame regions, the method attempts to preserve informative fire structure before centralization. A plausible implication is that the extraction stage is designed to avoid trivial oversampling of noisy or low-value fire fragments.

3. Centralization transformation and copy-paste operation

The centralization transformation is defined as a mechanism for extracting each fire cluster’s “core,” removing spurious boundary pixels (Kim et al., 8 Jul 2025). It applies morphological erosion with a small kernel $I_i \in \mathbb{R}^{H\times W\times 3}$4:

$I_i \in \mathbb{R}^{H\times W\times 3}$5

where $I_i \in \mathbb{R}^{H\times W\times 3}$6 denotes erosion and $I_i \in \mathbb{R}^{H\times W\times 3}$7 is an $I_i \in \mathbb{R}^{H\times W\times 3}$8 square, with $I_i \in \mathbb{R}^{H\times W\times 3}$9 recommended.

For each $M_i \in \{0,\dots,C-1\}^{H\times W}$0, the centroid in continuous coordinates is computed as

$M_i \in \{0,\dots,C-1\}^{H\times W}$1

If $M_i \in \{0,\dots,C-1\}^{H\times W}$2 is the minimal axis-aligned bounding rectangle of $M_i \in \{0,\dots,C-1\}^{H\times W}$3, the mask and corresponding image patch are shifted by

$M_i \in \{0,\dots,C-1\}^{H\times W}$4

where $M_i \in \{0,\dots,C-1\}^{H\times W}$5. Any pixels that would move outside the patch are clamped. The stated purpose is to re-center the bounding box of each $M_i \in \{0,\dots,C-1\}^{H\times W}$6 so that $M_i \in \{0,\dots,C-1\}^{H\times W}$7 becomes its geometric center.

Once a library of centralized cores $M_i \in \{0,\dots,C-1\}^{H\times W}$8 has been formed, augmented pairs $M_i \in \{0,\dots,C-1\}^{H\times W}$9 are generated by selecting a target image $C$0 and target mask $C$1, drawing a random core, and searching for a valid placement region within image bounds. Placement is accepted when the fraction of the candidate region already labeled background-only satisfies the overlap constraint governed by $C$2. With $C$3, the method ensures no overlap with existing fire. If placement succeeds, pixels under the pasted core are copied into the target image, optionally with boundary smoothing controlled by $C$4, and the corresponding target-mask pixels are set to fire.

The paper reports several heuristics: $C$5 copies per image, $C$6 to avoid infinite loops, $C$7 as the default overlap threshold, and optional rotation of the core by $C$8 to further increase diversity. The discussion attributes CCPDA’s advantage to “isolating the highest-confidence fire pixels (the ‘core’), which reduces label noise,” and to centering them so that convolutional receptive fields see flame patterns in their canonical spatial context.

4. Hyperparameters and recommended settings

The method description provides explicit recommendations for the principal hyperparameters (Kim et al., 8 Jul 2025).

Component	Setting	Stated role
$C$9	$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$0	Discards tiny noise clusters
$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$1	$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$2	Captures halo around flame
$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$3	$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$4	Yields stable “core”
$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$5	$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$6	Number of cores pasted per image
$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$7	$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$8	No overlap with existing fire
$F_i = \mathbf{1}_{[M_i==\mathrm{fire}]}.$9	$F_i$0	Limits placement retries

Within this range, the reported empirical recommendation is that $F_i$1 strikes a good balance. The overlap threshold can be relaxed to $F_i$2 for scenes with large fire. The blending flag $F_i$3 is described as controlling whether to smooth the boundary, for example by a linear alpha-blend of width $F_i$4.

These settings indicate that CCPDA is intended as a lightweight, plug-and-play augmentation rather than a heavily parameterized synthesis framework. This suggests that its design target is practical integration with existing training pipelines under small-sample conditions.

5. Weighted sum-based multi-objective evaluation

The evaluation protocol is explicitly fire-centric. Because fire segmentation has multiple operationally relevant metrics, methods are compared via a single scalar score

$F_i$5

subject to

$F_i$6

The reported experimental weights are

$F_i$7

Each metric is computed on a held-out test set, and higher $F_i$8 denotes better overall fire performance (Kim et al., 8 Jul 2025). This weighting scheme formalizes the paper’s central premise: improvement on the fire class is the dominant operational criterion, and a single aggregate score is used to compare augmentation strategies under that premise.

A common misconception would be to treat the method as optimizing generic semantic-segmentation quality in a class-agnostic way. The stated evaluation does not do that. It aggregates $F_i$9, $\{s_{i,j}\}$0, $\{s_{i,j}\}$1, and $\{s_{i,j}\}$2, and therefore prioritizes fire-class behavior by construction.

6. Numerical results, limitations, and future directions

The reported numerical results summarize mean $\{s_{i,j}\}$3 $\{s_{i,j}\}$4 std over $\{s_{i,j}\}$5 random splits for the fire class only (Kim et al., 8 Jul 2025). Baseline Aug (flip,rot) achieved $\{s_{i,j}\}$6, $\{s_{i,j}\}$7, $\{s_{i,j}\}$8, $\{s_{i,j}\}$9, and weighted $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$0. Standard Copy-Paste achieved $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$1, $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$2, $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$3, $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$4, and $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$5. CCPDA achieved $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$6, $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$7, $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$8, $$S_i = \{s_{i,j} \subseteq \Omega \mid s_{i,j} \text{ is a connected set of pixels in } F_i \text{ and } \Omega \text{ is the image domain}\}.$$9, and $s'_{i,j} = s_{i,j} \oplus K_5,$0.

The statistical-significance result reported for the comparison between CCPDA and Standard Copy-Paste is $s'_{i,j} = s_{i,j} \oplus K_5,$1 for $s'_{i,j} = s_{i,j} \oplus K_5,$2 and $s'_{i,j} = s_{i,j} \oplus K_5,$3. The paper therefore states that CCPDA outperforms other augmentation strategies in the application scenario considered, particularly in improving fire-class segmentation performance.

The paper also states several limitations. Pasted cores may violate physical constraints, for example fire floating above the ground. There is no dynamic lighting adaptation, so the color and brightness of pasted cores sometimes clash with target scenes. Fixed erosion and dilation kernels may under- or over-shrink clusters in extremely large or tiny fires. These points qualify a possible misunderstanding that copy-paste realism is guaranteed by centralization alone; the authors do not claim that.

Future extensions proposed in the paper include context-aware placement via depth or ground-plane estimation, learned alpha-matte generation for physically plausible blending, adaptive kernel sizes based on cluster geometry, and integration with generative adversarial augmentation (GAN-CCPDA). In that sense, CCPDA is presented not as a complete physical simulation model, but as a lightweight augmentation framework whose empirical value lies in focusing on genuine flame cores, recentralizing them, and carefully pasting them into novel backgrounds.