Wavelet-Guided Dual-Frequency Encoding

• The paper introduces WGDF, a dual-branch framework using discrete wavelet transform to separate detailed edge information from global semantic context for improved remote sensing change detection.
• It employs high-frequency modules (DFFE and FDID) to enhance local edge and texture changes and low-frequency modules (Transformers and PCDM) to capture scene-level structures.
• Evaluation on LEVIR-CD, WHU-CD, and GZ-CD datasets shows notable gains in F1, IoU, and Edge mIoU while maintaining efficient computational performance.

Wavelet-Guided Dual-Frequency Encoding (WGDF) is a dual-branch framework for remote sensing change detection in which Discrete Wavelet Transform (DWT) is used at the input stage to separate high-frequency and low-frequency components, allowing local details and global structures to be modeled by different computational pathways. As introduced in "Wavelet-Guided Dual-Frequency Encoding for Remote Sensing Change Detection," the method is motivated by the claim that spatial-domain feature modeling has limited diversity for subtle change perception, whereas wavelet-domain decomposition can better expose edge changes, fine textures, and coarse semantic layout; WGDF therefore combines a high-frequency branch built around Dual-Frequency Feature Enhancement (DFFE) and Frequency-Domain Interactive Difference (FDID) with a low-frequency branch built around Transformer blocks and a Progressive Contextual Difference Module (PCDM), followed by synergistic fusion and inverse wavelet reconstruction (Zhang et al., 7 Aug 2025).

1. Problem setting and frequency-domain rationale

WGDF is situated in remote sensing change detection, especially building change detection, where the target signal is often weak, spatially localized, and easily obscured by clutter. The method description identifies several recurring difficulties: small buildings, narrow edges, partially occluded structures, and low-contrast changes are easy to miss; backgrounds such as vegetation, roads, shadows, water, and textured surfaces introduce strong interference; and purely spatial-domain models tend to entangle semantics, texture, and edges, which in turn leads to weak sensitivity to tiny structural changes, blurred boundaries, false positives in cluttered regions, and missed detections of small objects (Zhang et al., 7 Aug 2025).

The central rationale of WGDF is that wavelet decomposition permits an explicit division of labor between two spectral regimes. Low-frequency information is assigned to global semantics, coarse structures, and contextual layout, whereas high-frequency information is assigned to edges, contours, fine textures, and localized variations. This decomposition is not treated as a late auxiliary operation; rather, the paper emphasizes that the input is decomposed “at the source” using DWT. A plausible implication is that the model attempts to reduce representational interference before multi-scale reasoning and differencing are applied.

The method uses Haar DWT. In the formulation given for a bi-temporal image pair $x_i$ and $y_i$,

$DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$

$DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$

Here $LL$ denotes the low-frequency approximation, while $LH$, $HL$, and $HH$ denote high-frequency detail bands. The paper characterizes Haar DWT as computationally efficient and as preserving both spatial and frequency locality (Zhang et al., 7 Aug 2025).

2. Architectural organization and dual-frequency decomposition

WGDF is a dual-branch network whose workflow is organized around wavelet decomposition, branch-specific modeling, and reconstruction. After DWT, the high-frequency branch receives the three detail bands and targets building edges, small structural changes, contours, subtle local differences, and high-frequency temporal variation. The low-frequency branch receives the approximation band and targets global semantic context, object-level structure, scene layout, and semantic change regions in complex backgrounds. The two outputs are fused and reconstructed through inverse DWT (IDWT), after which a classifier produces the final change prediction (Zhang et al., 7 Aug 2025).

The overall design reflects a strict separation between local sensitivity and global discriminability. In the high-frequency path, the method emphasizes enhancement and differencing of subtle local changes. In the low-frequency path, it emphasizes contextual reasoning and progressive semantic refinement. The paper states that the final high- and low-frequency features are “synergistically fused” to combine local sensitivity from the high-frequency branch with global discriminability from the low-frequency branch (Zhang et al., 7 Aug 2025).

The reconstruction step is given as

$$Change\ map = IDWT(D^{HH}, D^{HL}, D^{LH}, D^{LL}),$$

where $D^{HH}$, $y_i$0, and $y_i$1 are high-frequency difference features and $y_i$2 is the low-frequency difference feature. This formulation is notable because it keeps the change map in explicit correspondence with wavelet subbands until the final stage, rather than collapsing frequency-separated features prematurely (Zhang et al., 7 Aug 2025).

A common misconception is that WGDF is merely wavelet preprocessing followed by a conventional detector. The module definitions and ablations argue against that interpretation: the reported gains are attributed not only to DWT itself but to the branch-specific enhancement and difference modeling implemented by DFFE, FDID, Transformer blocks, and PCDM (Zhang et al., 7 Aug 2025).

3. High-frequency branch: DFFE and FDID

The high-frequency branch is designed for edge detail representation and fine-grained change modeling. Its first dedicated component is Dual-Frequency Feature Enhancement (DFFE), which strengthens high-frequency detail representation while suppressing noise and redundant background responses. DFFE receives high-frequency features from both time steps, denoted $y_i$3 and $y_i$4, computes a difference cue $y_i$5, and then applies attention-style recalibration (Zhang et al., 7 Aug 2025).

The paper gives the following formulation: $y_i$6

$y_i$7

In this design, $y_i$8 compresses channels, max and average pooling provide complementary descriptors, and $y_i$9 produces an attention map that reweights the original high-frequency features. The stated purpose is to boost discriminative edge responses while suppressing noise (Zhang et al., 7 Aug 2025).

DFFE is stacked three times. The paper reports that three stacked DFFE blocks provide the best balance; fewer blocks underfit, whereas too many may introduce redundancy or instability. This supports the interpretation that the high-frequency branch is intended to be deep enough for refinement but not so deep that detail cues are over-processed (Zhang et al., 7 Aug 2025).

The second major component is the Frequency-Domain Interactive Difference (FDID) module, which models fine-grained high-frequency change differences between the two times while remaining robust to texture noise and illumination variation. FDID begins with an absolute high-frequency difference: $DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$0 It then applies three parallel convolutions with kernel sizes $DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$1, $DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$2, and $DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$3, each followed by CBAM: $DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$4

$DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$5

After this multi-scale aggregation, channel attention is applied through global average pooling and an MLP: $DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$6

$DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$7

The stated effect of FDID is multi-scale difference extraction, attention-based suppression of irrelevant background responses, better localization of subtle edge changes, and stronger fine-grained temporal modeling in high-frequency space (Zhang et al., 7 Aug 2025).

4. Low-frequency branch: Transformer modeling and PCDM

The low-frequency branch processes the $DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$8 bands and is responsible for global semantic context, object-level structure, and scene layout. The method places stacked Transformer blocks in this branch because the $DWT(x_i) = [HH_1, HL_1, LH_1, LL_1]$9 representation is where long-range dependencies and scene-level semantic relationships can be captured most directly. The paper describes the workflow as taking $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$0 and $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$1 and refining them through repeated Transformer-based interaction: $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$2 The notation is compressed in the original description, but the intended role is explicit: repeated Transformer reasoning over low-frequency semantic features (Zhang et al., 7 Aug 2025).

The Progressive Contextual Difference Module (PCDM) is the low-frequency counterpart to FDID. Its purpose is to refine low-frequency semantic differences progressively and improve localization in cluttered backgrounds. PCDM begins from a low-frequency absolute difference: $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$3 That difference cue is then reinjected into the original low-frequency features: $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$4 This is followed by a progressive dilated-convolution sequence with dilation rates $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$5, $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$6, $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$7, and $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$8: $DWT(y_i) = [HH_2, HL_2, LH_2, LL_2].$9

$LL$0

$LL$1

$LL$2

Finally, the low-frequency difference representation is formed as

$LL$3

where $LL$4 denotes concatenation (Zhang et al., 7 Aug 2025).

The paper attributes several functions to PCDM: coarse-to-fine contextual differencing, progressive sharpening of semantic change regions, suppression of interference from vegetation and roads, and improved change completeness. A plausible interpretation is that PCDM operationalizes the idea that low-frequency information should not be treated as merely smooth background; instead, it is the carrier of structural semantics and context needed to reject false alarms that arise from high-frequency clutter alone.

5. Fusion, optimization, datasets, and empirical behavior

The loss function combines BCE loss and Dice loss to address class imbalance: $LL$5

$LL$6

$LL$7

The best setting reported is $LL$8 and $LL$9. The paper states that BCE improves pixel-wise classification accuracy, Dice improves region overlap under imbalance, and the combination reduces edge ambiguity while improving completeness (Zhang et al., 7 Aug 2025).

Training is reported in concrete implementation terms: PyTorch 1.11; 2 $LH$0 NVIDIA TITAN RTX; AdamW; momentum $LH$1; weight decay $LH$2; $LH$3; $LH$4 epochs; learning rate $LH$5; batch size $LH$6; and augmentations including flipping, scaling, cropping, and Gaussian blur. The benchmark datasets are LEVIR-CD, WHU-CD, and GZ-CD. Evaluation uses Precision, Recall, F1-score, IoU, Overall Accuracy, and Edge mIoU, where

$LH$7

$LH$8

The emphasis on Edge mIoU is consistent with the method’s stated objective of reducing edge ambiguity (Zhang et al., 7 Aug 2025).

WGDF is compared against ten methods: BIT, HFA-Net, ChangeFormer, FTN, ICIF-Net, AERNet, ELGC-Net, SEIFNet, CF-GCN, and ChangeBind. The reported results identify WGDF as best or near-best on all three datasets, with particular strength in F1, IoU, and Edge mIoU. On LEVIR-CD, the best reported values are F1 $LH$9, IoU $HL$0, and Edge mIoU $HL$1, exceeding ChangeBind by $HL$2 F1, $HL$3 IoU, and $HL$4 Edge mIoU. On WHU-CD, the reported best values are F1 $HL$5, IoU $HL$6, Recall $HL$7, and Edge mIoU $HL$8. On GZ-CD, the reported best values are F1 $HL$9, IoU $HH$0, and Edge mIoU $HH$1. The paper also reports moderate complexity at approximately $HH$2M parameters, lower FLOPs than FTN and ChangeFormer, and faster inference than those heavier models (Zhang et al., 7 Aug 2025).

The ablations are central to the interpretation of WGDF. The best depth is reported as three DFFE blocks and three Transformer blocks. Removing FDID or PCDM reduces performance. BCE alone or Dice alone is weaker than the combined loss. These findings support the claim that WGDF’s gains arise from the division of labor between branches and the specialized refinement modules, rather than from wavelet decomposition alone (Zhang et al., 7 Aug 2025).

6. Relation to adjacent wavelet-guided frequency methods

WGDF belongs to a broader class of frequency-aware architectures that use wavelets to separate or regulate spectral content, but the precise architectural meaning of “wavelet-guided” varies across tasks. In retinal vessel segmentation, WaveRNet frames domain shift as frequency-dependent, separates illumination-stable low-frequency structures from high-frequency vessel boundaries through a Spectral-guided Domain Modulator, performs test-time frequency-based domain fusion, and adds a Hierarchical Mask-Prompt Refiner for detail recovery inside a SAM-based framework (Wang et al., 9 Jan 2026). In semantic segmentation, WaveSeg similarly decomposes features into $HH$3, $HH$4, $HH$5, and $HH$6, applies re-parameterized convolutions to preserve low-frequency semantic integrity, applies Mamba/VSSM to high-frequency subbands, and fuses boundary-aware and semantic representations in a decoder-centric design (Xu et al., 24 Oct 2025).

In restoration and generation, the same dual-frequency intuition appears with different operational choices. SPJFNet uses 2D-DWT to generate low- and high-frequency pathways, then restores low-frequency content in the Fourier domain and enhances high-frequency wavelet detail with gradient-guided gating; its reported decomposition into a DFGF low-frequency branch and DFGF high-frequency branch makes it strongly WGDF-like, though it is formulated for dark image restoration and emphasizes efficiency and self-mined priors (Zhang et al., 6 Aug 2025). HDW-SR uses wavelet-based downsampling inside a residual diffusion network, where low-frequency diffused features query high-frequency prior features from a pre-super-resolved image through sparse cross-attention; the paper explicitly interprets the design as separating structure from detail and recombining them with low-loss inverse wavelet reconstruction (Yang et al., 17 Nov 2025). Zero-shot low-light enhancement via joint wavelet and Fourier priors also relies on the proposition that illumination is concentrated in low-frequency components and structure in high-frequency components, but it implements guidance as iterative sampling-time recombination in a pre-trained diffusion model rather than as a dedicated dual-branch encoder (He et al., 2024).

Not every wavelet-guided method is dual-frequency in the strict WGDF sense. ImplicitTerrainV2 uses a wavelet complexity field derived from stationary wavelet coefficients and gradient magnitude to spatially gate multiple frequency bands within a single SIREN backbone, and the paper explicitly distinguishes this from a classic dual-branch low-frequency/high-frequency encoding architecture (Feng et al., 21 May 2026). This contrast clarifies an important terminological point: WGDF, as defined in remote sensing change detection, is not equivalent to any use of wavelets for spectral control. It refers more specifically to explicit low-/high-frequency decomposition, branch-specialized processing, and final synergistic fusion (Zhang et al., 7 Aug 2025).

The future direction stated for WGDF itself is further network refinement and lightweight design for larger and more complex scenarios (Zhang et al., 7 Aug 2025). A plausible implication, supported by the neighboring literature, is that later WGDF-style systems may continue to hybridize wavelet decomposition with Transformers, Mamba-style sequence modeling, diffusion priors, or domain-adaptive inference while preserving the core principle that local detail and global structure are best handled in distinct but coupled frequency pathways.