CGCCE-Net: Enhanced Cross-Correlation for Change Detection

• The paper presents CGCCE-Net, a PVT-based siamese encoder–decoder that improves detection of special-colored buildings in remote sensing change tasks.
• It integrates texture-guided refinement, global cross-temporal semantic correlation, and multi-scale semantic enhancement to produce accurate pixel-level building change maps.
• Empirical results on LEVIR-CD, WHU-CD, and GZ-CD show notable improvements in F1 and IoU compared to state-of-the-art change detection methods.

Searching arXiv for the specified paper and closely related change detection work to ground the article in current research. Change-Guided Cross Correlation Enhancement Network (CGCCE-Net) is a remote sensing building change detection model formulated as a PVT-based, siamese encoder–decoder network that integrates texture-guided shallow refinement, cross-temporal semantic correlation, semantic cognitive enhancement, and cross-fusion decoding. It is designed for the standard bi-temporal setting in which two co-registered remote sensing images of the same area, acquired at times $T_1$ and $T_2$, are used to predict a pixel-level binary change map of building changes. The method is explicitly motivated by failure modes in building change detection involving special-color buildings, subtle and small structural changes, noisy backgrounds, and complex dense scenes (Wang, 3 Aug 2025).

1. Problem formulation and design objective

In building change detection, the input is a pair of co-registered remote sensing images of the same area at two times $T_1$ and $T_2$, and the output is a pixel-level change map highlighting changed building regions. The target changes include buildings that have appeared, disappeared, or changed, and the paper frames the task as accurate localization plus discrimination of changed building areas (Wang, 3 Aug 2025).

The central motivation of CGCCE-Net is the observation that many existing deep learning-based building change detection methods, including U-Net variants, CNN-based models, and Transformer-based models, still exhibit three specific weaknesses. First, many networks learn a building color prior and fail on roofs with deep purple, bright white, or other unusual colors. Second, partial extensions, incomplete demolitions, narrow strips, and fine structures are easily missed. Third, vegetation, roads, shadows, seasonal variations, and illumination changes can be misclassified as building changes. The paper therefore targets a setting in which the challenge is not only generic temporal differencing, but robust recognition of buildings under color atypicality, background clutter, and cross-temporal appearance variation (Wang, 3 Aug 2025).

CGCCE-Net is intended to fill these gaps through three coordinated design choices: explicit guidance from shallow texture and color cues, specialized global cross-temporal semantic interaction, and multi-scale semantic enhancement before decoding. The explicit target is to significantly improve detection of buildings with special colors while maintaining or improving general building change detection performance. This suggests that the model is not merely a generic bi-temporal differencing architecture, but a task-specialized recognition system for building-focused remote sensing change analysis (Wang, 3 Aug 2025).

2. Architectural composition

The overall architecture is a siamese, PVT-based encoder–decoder. Two branches, corresponding to $T_1$ and $T_2$, share weights and extract shallow and deep multi-scale features: $T_{ij} = PVT_{\text{Encoder}}(T_i)$ where $i \in \{1,2\}$ indexes time and $j$ indexes scale (Wang, 3 Aug 2025).

The network is organized around six principal components.

Component	Inputs	Role
PVT encoder	$T_1, T_2$	Extracts shallow texture features and deep semantic features
CGRR branch	Shallow features $T_2$0	Extends shallow texture features to multiple scale features and guides early attention
CAB and SAB	Deep features	Perform independent information enhancement on each time branch
GCCM	Attention-enhanced deep features	Computes linear-angle-attention-based cross correlation between times
SCEM	Concatenated and CGRR-enhanced features $T_2$1	Combines multi-kernel local context and global context for semantic enhancement
CFD	Multi-scale enhanced features	Fuses change information and reconstructs the final binary change map

The data flow is sequential but interdependent. Inputs first pass through the PVT encoder to produce shallow and deep features for both times. Shallow features are processed by the change-guided residual refinement branch and fed back into concatenated representations $T_2$2. Deep features are independently enhanced by channel attention and spatial attention, then used by the global cross correlation module to generate semantically cross-correlated features $T_2$3 and $T_2$4. The semantically enhanced concatenated features from SCEM and the cross-temporal features from GCCM are finally integrated by the cross fusion decoder, which reconstructs the pixel-level change map (Wang, 3 Aug 2025).

This organization separates three kinds of information that are often conflated in simpler change detectors: local texture and color evidence, high-level semantic correspondence across time, and decoder-stage reconstruction of change boundaries. A plausible implication is that the architecture is designed to reduce the common trade-off between fine local sensitivity and robust global discrimination.

3. Change-guided residual refinement branch

The Change-Guided Residual Refinement (CGRR) branch operates on shallow PVT features and addresses missed detections of special-colored buildings by exploiting low-level texture and color details. The paper defines two basic operations over multi-scale features: $T_2$5

$T_2$6

where $T_2$7 captures raw change response and $T_2$8 preserves joint bi-temporal information (Wang, 3 Aug 2025).

CGRR uses multi-convolutional layers and a residual structure to refine shallow features: $T_2$9 Here, $T_1$0 denotes stacked convolutional blocks such as Conv–BN–ReLU, and $T_1$1 denotes a residual wrapper that preserves the original signal while adding refined cues. Conceptually, the branch learns texture feature maps emphasizing special roof colors, edges, and local intensity transitions, then feeds these cues back into $T_1$2 so that later modules receive texture-aware change information rather than only high-level semantic abstractions (Wang, 3 Aug 2025).

The branch is significant because it shifts the burden of recognizing atypical roofs away from purely deep semantics. In ordinary encoder–decoder designs, unusual roof colors may be suppressed by learned class regularities; in CGCCE-Net, shallow texture and color evidence is explicitly preserved and reinjected. The ablation study further supports this role: removing CGRR weakens detection on special-color buildings, especially in the qualitative sample identified as sample F, where the model becomes specifically weak on such targets (Wang, 3 Aug 2025).

4. Cross-temporal semantic correlation and semantic cognition

Before cross-temporal interaction, deep features undergo independent information enhancement using channel attention and spatial attention. Channel attention learns per-channel weights through global pooling and an MLP in a typical Squeeze-and-Excitation style, emphasizing building-related channels and suppressing channels associated with vegetation, roads, or illumination artifacts. Spatial attention learns a spatial weighting map that highlights locations likely to contain buildings and potential changes. The purpose of this stage is to provide cleaner, building-focused deep representations for subsequent correlation learning (Wang, 3 Aug 2025).

The Global Cross Correlation Module (GCCM) is the principal cross-temporal interaction mechanism. It computes queries, keys, and values from attention-enhanced deep features and uses a linear-angle similarity approximation: $T_1$3 This formulation approximates normalized angular similarity while avoiding the expensive $T_1$4 operation. The resulting cross-correlation features are

$T_1$5

and

$T_1$6

for times 1 and 2, respectively (Wang, 3 Aug 2025).

GCCM differs from simple elementwise differencing or raw correlation because it uses a learned Q/K/V attention framework and an angle-based linear similarity approximation to encode semantic recognition relationships between bi-temporal features. In the building change detection setting, this is intended to establish how a building at $T_1$7 corresponds to structures at $T_1$8, thereby improving robustness against illumination changes and seasonal effects rather than merely amplifying appearance differences (Wang, 3 Aug 2025).

The Semantic Cognitive Enhancement Module (SCEM) further refines the concatenated feature $T_1$9 by combining local multi-scale receptive fields and global context: $T_2$0

$T_2$1

$T_2$2

This mechanism reallocates feature weights through multi-kernel local processing and globally pooled semantic gating, enhancing long-range dependencies and semantic recognition relationships for bi-temporal features (Wang, 3 Aug 2025).

The Cross Fusion Decoder (CFD) integrates the outputs of CGRR, GCCM, and SCEM. Its stated role is to perform feature interaction between branches through cross attention, multi-scale fusion of enhanced encoder outputs, and spatial resolution recovery to the original image size. The output is a final binary change map. This suggests that CGCCE-Net’s decoder is not a passive reconstruction head; it is the stage where texture-guided cues, cross-temporal semantics, and semantically reweighted multi-scale features are reconciled into a single pixel-level decision function (Wang, 3 Aug 2025).

5. Optimization, datasets, and evaluation protocol

CGCCE-Net is trained with binary cross-entropy on the final change map: $T_2$3 where $T_2$4 is the number of pixels, $T_2$5 is the ground-truth label, and $T_2$6 is the predicted probability. No auxiliary loss is mentioned; optimization is based purely on the final change map (Wang, 3 Aug 2025).

The implementation uses PyTorch on Ubuntu 18.04 with an NVIDIA TITAN RTX 24GB. Training runs for 500 epochs, with validation at each epoch and best-model saving. The optimizer is AdamW, the initial learning rate is $T_2$7, and the learning rate schedule is cosine annealing. All datasets are cropped into $T_2$8 patches, and random cropping is explicitly used to generate patches. A learning-rate sensitivity analysis reports stable performance for learning rates in $T_2$9, best F1 and IoU at $T_1$0, slow and suboptimal convergence at $T_1$1, and unstable non-convergent training at $T_1$2 (Wang, 3 Aug 2025).

The model is evaluated on LEVIR-CD, WHU-CD, and GZ-CD. LEVIR-CD contains 637 image pairs of size $T_1$3, cropped to $T_1$4, with a split of 70% training, 20% testing, and 10% validation. WHU-CD is cropped to $T_1$5 for a total of 7620 pairs, with 80% training, 10% testing, and 10% validation. GZ-CD, after cropping to $T_1$6, contains 2834 training pairs, 325 testing pairs, and 400 validation pairs. Evaluation uses F1-score, IoU, Precision, and Recall, computed from pixel-level TP, FP, and FN (Wang, 3 Aug 2025).

6. Empirical performance and ablation evidence

CGCCE-Net is compared against FC-Siam-Conc, IFNet, SNUNet, BIT, DTCDSCN, ChangeFormer, ICIF-Net, DMINet, AERNet, and SEIFNet. On LEVIR-CD, it achieves $T_1$7, $T_1$8, $T_1$9, and $T_2$0. The paper identifies ICIF-Net and SEIFNet as the next-best models on F1 and IoU, and reports improvements over SEIFNet of $T_2$1 F1 and $T_2$2 IoU. On WHU-CD, CGCCE-Net achieves $T_2$3, $T_2$4, $T_2$5, and $T_2$6, improving over SEIFNet by $T_2$7 F1 and $T_2$8 IoU. On GZ-CD, it reaches $T_2$9, $T_{ij} = PVT_{\text{Encoder}}(T_i)$0, $T_{ij} = PVT_{\text{Encoder}}(T_i)$1, and $T_{ij} = PVT_{\text{Encoder}}(T_i)$2, improving over SEIFNet by $T_{ij} = PVT_{\text{Encoder}}(T_i)$3 F1 and $T_{ij} = PVT_{\text{Encoder}}(T_i)$4 IoU (Wang, 3 Aug 2025).

The qualitative analysis is tightly aligned with the model’s stated design goals. On LEVIR-CD, CGCCE-Net avoids false positives on construction sites, preserves fine structure when many methods merge two rows of small buildings, and produces more complete predictions for buildings with alternating shadow and white colors. On WHU-CD, it detects less obvious color changes, uniquely captures a full building missed by other methods under subtle color variation, and avoids redundant predictions in large change regions mixed with roads. On GZ-CD, it detects true changes while discarding false ones in scenes with similar-color change regions, and it performs best on deep blue or other special-colored buildings, with sharper edges and fewer misses. Figure 1 is singled out as showing successful detection of deep purple, bright white, and otherwise atypical roofs (Wang, 3 Aug 2025).

A real Google Earth scene from Oxford, Mississippi, covering 2012–2019 and lacking ground truth, is also presented. The paper states that, visually, CGCCE-Net yields more structurally organized building change maps with cleaner edges and more plausible building outlines than the baselines. Because no ground truth is available, this should be interpreted as qualitative evidence of practicality rather than a benchmarked accuracy claim (Wang, 3 Aug 2025).

The ablation study removes GCCM, CGRR, SCEM, and CFD in turn. Removing any component degrades F1 and IoU on all datasets, and the full model consistently outperforms every ablated version. Qualitative ablation results indicate that without GCCM the model misses small buildings and produces intertwined or misaligned predictions; without CGRR it becomes specifically weak on special-color buildings; and without SCEM or CFD the outputs show less precise edges and weaker global consistency. The ablation evidence therefore supports a modular interpretation of the architecture: CGRR supplies texture-guided local evidence, GCCM supplies cross-temporal semantic correspondence, SCEM improves semantic reweighting, and CFD consolidates these cues into final predictions (Wang, 3 Aug 2025).

7. Computational profile, limitations, and position in the literature

CGCCE-Net has 56.67M parameters and 17.57G FLOPs. The paper notes that some models, such as ChangeFormer, have lower parameter count but much higher computational cost, citing 41.03M parameters and 202.83G FLOPs for ChangeFormer. It also notes that some alternatives, including BIT and SEIFNet, are lighter in parameters but deliver lower performance. The interpretation offered is that the model’s PVT-based siamese encoder together with CGRR and GCCM increases parameter count, while the FLOPs remain relatively low for the achieved performance (Wang, 3 Aug 2025).

The authors explicitly identify several limitations. CGRR introduces local noise and increases computational overhead. GCCM’s global semantic interaction remains susceptible to noise in long-range dependencies, which can degrade global feature robustness. The network lacks a more refined transition mechanism and fusion strategy for change information. Figure 2 is reported to show both missing target predictions and redundant predictions in challenging scenes. The method is also specialized for building changes, which may limit generalization to broader land-cover change detection tasks (Wang, 3 Aug 2025).

Future work is framed in terms of more balanced treatment of general remote sensing change detection challenges and specific problems such as special-color building changes, improved noise robustness in both local and global representations, better fusion and transition strategies for change information, and extension to richer and more diverse change scenarios beyond buildings (Wang, 3 Aug 2025).

Within the broader change-detection literature, CGCCE-Net occupies a position between generic Transformer-based context modeling and explicitly relationship-aware bi-temporal modeling. Relative to CNN-based methods such as FC-Siam-Conc, IFNet, and SNUNet, it retains local convolutional processing in CGRR but introduces a PVT backbone and explicit global cross-temporal interaction. Relative to attention-oriented designs, its GCCM is specialized for bi-temporal semantic cross-correlation in remote sensing building change detection rather than generic attention alone (Wang, 3 Aug 2025). Related work further clarifies this placement. SRC-Net emphasizes reliability-aware bi-temporal spatial relationship modeling through cross-branch perception and patch-mode fusion (Chen et al., 2024), while CGNet emphasizes change-prior-guided multi-scale self-attention for edge integrity and hole reduction (Han et al., 2024). This suggests that CGCCE-Net can be understood as an overview-oriented architecture in which shallow texture guidance, cross-temporal semantic correlation, and semantic reweighting are jointly optimized for remote sensing building change detection.