PaperlayerView exact revision on arXiv

Source-linked AI summary

FSG-Net: Frequency-Spatial Synergistic Gated Network for High-Resolution Remote Sensing Change Detection

Zhongxiang Xie, Shuangxi Miao, Yuhan Jiang, Zhewei Zhang, Jing Yao, Xuecao Li, Jianxi Huang, Pedram Ghamisi

arXiv:2509.06482v1cs.CV

TL;DR

High-resolution remote-sensing change detection must separate semantic changes from imaging-condition variation and align deep features with shallow detail. FSG-Net combines wavelet-based frequency processing, spatial attention, and semantics-guided gated fusion; it reports state-of-the-art performance on three benchmarks, with F1-scores of 94.16% on CDD, 89.51% on GZ-CD, and 91.27% on LEVIR-CD.

  • Problem

    Illumination and seasonal variation can appear as pseudo-changes, while the semantic gap between deep abstract features and shallow details hinders precise boundary delineation.

    Source: §I · p. 1

  • Method

    FSG-Net combines frequency-domain wavelet interaction, spatial-temporal attention, and deep-semantics-driven gating of shallow features.

    Source: §I · p. 2

  • Results

    FSG-Net reports state-of-the-art F1-scores of 94.16% on CDD, 89.51% on GZ-CD, and 91.27% on LEVIR-CD.

    Source: F. Discussion · p. 12

  • Takeaways & Limitations

    The reported results and visualizations indicate that the frequency–spatial interplay suppresses pseudo-changes and enhances genuine changes across the evaluated datasets.

    Source: F. Discussion · p. 12

Abstract

from arXiv · show

Change detection from high-resolution remote sensing images lies as a cornerstone of Earth observation applications, yet its efficacy is often compromised by two critical challenges. First, false alarms are prevalent as models misinterpret radiometric variations from temporal shifts (e.g., illumination, season) as genuine changes. Second, a non-negligible semantic gap between deep abstract features and shallow detail-rich features tends to obstruct their effective fusion, culminating in poorly delineated boundaries. To step further in addressing these issues, we propose the Frequency-Spatial Synergistic Gated Network (FSG-Net), a novel paradigm that aims to systematically disentangle semantic changes from nuisance variations. Specifically, FSG-Net first operates in the frequency domain, where a Discrepancy-Aware Wavelet Interaction Module (DAWIM) adaptively mitigates pseudo-changes by discerningly processing different frequency components. Subsequently, the refined features are enhanced in the spatial domain by a Synergistic Temporal-Spatial Attention Module (STSAM), which amplifies the saliency of genuine change regions. To finally bridge the semantic gap, a Lightweight Gated Fusion Unit (LGFU) leverages high-level semantics to selectively gate and integrate crucial details from shallow layers. Comprehensive experiments on the CDD, GZ-CD, and LEVIR-CD benchmarks validate the superiority of FSG-Net, establishing a new state-of-the-art with F1-scores of 94.16%, 89.51%, and 91.27%, respectively. The code will be made available at https://github.com/zxXie-Air/FSG-Net after a possible publication.

I. INTRODUCTION

High-resolution remote-sensing change detection must distinguish semantic changes from acquisition-related pseudo-changes while aligning deep semantics with shallow details. FSG-Net addresses these challenges through frequency–spatial interaction and semantics-guided feature fusion.

I. INTRODUCTION

  • Illumination and seasonal shifts create pseudo-changes, while the gap between deep concepts and shallow details degrades change-boundary precision.The task is pixel-wise identification of semantic changes in co-registered images acquired at different times.Source: §I · p. 1
  • Existing spatial and frequency-domain approaches leave room for improvement in suppressing background noise and jointly using frequency- and spatial-domain interactions.Prior multi-level fusion and alignment mechanisms can also incur high computational complexity.Source: §I · p. 2
  • FSG-Net combines frequency–spatial separation to distinguish changes from nuisance variations with deep semantics that refine multi-level feature details.The approach targets both pseudo-change suppression and semantics-aware alignment and fusion.Source: §I · pp. 1–2
  • DAWIM adaptively processes frequency sub-bands to mitigate radiometric-shift false alarms, STSAM amplifies genuine-change saliency, and LGFU uses semantic gates to fuse shallow details.LGFU is designed to yield crisp boundaries with minimal computational overhead.Source: §I · p. 2

II. RELATED WORK

Deep learning change-detection research spans CNN, Transformer, and hybrid architectures, each with different strengths and limitations. Bi-temporal feature interaction is central to change detection, and the paper identifies frequency–spatial interaction as a way to combine complementary forms of processing.

A. Deep Learning-Based Change Detection

  • CNNs capture contextual information but lack long-range dependency modeling, while Transformers model long-range relationships yet can be data-hungry, computationally demanding, and weak at fine boundaries.Hybrid architectures combine CNN features with Transformer context; FSG-Net incorporates enhanced CrossAtt into a CNN architecture.Source: A. Deep Learning-Based Change Detection · p. 3

B. Bi-temporal Feature Interaction in Change Detection

  • Bi-temporal feature interaction is important for change discrimination, but spatial interaction methods leave semantic content and stylistic variation entangled.Prior work explored joint attention, co-attention, and multi-level interaction.Source: B. Bi-temporal Feature Interaction in… · p. 3
  • Frequency-domain interaction methods have been explored, but existing work has insufficiently examined collaboration between frequency- and spatial-domain interactions.The paper proposes using frequency interaction to suppress pseudo-changes and spatial interaction to enhance genuine changes.Source: B. Bi-temporal Feature Interaction in… · p. 4

III. METHODOLOGY

FSG-Net uses a staged pipeline: DAWIM suppresses pseudo-changes in the frequency domain, STSAM enhances genuine-change saliency spatially, and LGFU harmonizes multi-level features for boundary delineation. DAWIM applies component-specific wavelet processing and adaptively weights the resulting interactions.

A. Method Overview

  • FSG-Net cascades DAWIM for frequency-domain pseudo-change suppression, STSAM for spatial enhancement of genuine changes, and LGFU for multi-level boundary refinement.The stages are presented as a pipeline addressing both task-irrelevant interference and the semantic gap.Source: A. Method Overview · p. 4

B. Discrepancy-Aware Wavelet Interaction Module (DAWIM)

  • DAWIM tailors processing to frequency components, using wavelet analysis for its time-frequency localization and multi-frequency sub-band analysis.Its workflow is shown in Fig. 2.Source: B. Discrepancy-Aware Wavelet… · p. 4
  • A 2D Haar DWT decomposes each bi-temporal feature map into LL, LH, HL, and HH subcomponents for component-specific processing.The decomposition uses predefined wavelet filters followed by downsampling.Source: B. Discrepancy-Aware Wavelet… · p. 4
  • Low-frequency LL components capture large-scale patterns and gradual radiometric variation; concatenation across times and 3D convolution model their temporal dynamics.The 3D convolution uses a 2×3×3 kernel to fuse temporal features and capture local spatial context.Source: B. Discrepancy-Aware Wavelet… · p. 4
  • Mid-frequency LH and HL components represent localized temporal variations and are fused with 3D convolution using a 2×1×1 kernel.The kernel fuses across time without compromising spatial structure.Source: B. Discrepancy-Aware Wavelet… · p. 5
  • High-frequency HH components represent edges, contours, and fine structural details; DAWIM computes their temporal difference feature.The cited equation expresses the difference using element-wise subtraction.Source: B. Discrepancy-Aware Wavelet… · p. 5
  • Modified SENet-based adaptive weights use max and average pooling to refine original frequency components with bi-temporal interaction features, after which IDWT reconstructs spatial-domain maps.The weighting is intended to enhance change-region saliency while mitigating irrelevant background noise.Source: B. Discrepancy-Aware Wavelet… · p. 5

C. Synergistic Temporal-Spatial Attention Module (STSAM)

STSAM combines temporal CrossAtt and spatial CoordAtt to model relationships between the two image times and refine local structure. Its two branches support global temporal interaction and fine-grained spatial attention.

C. Synergistic Temporal-Spatial Attention Module (STSAM)

  • STSAM pairs CrossAtt with CoordAtt for global temporal interaction and local spatial refinement.CrossAtt processes both temporal instances, while CoordAtt refines local structure, including fine-grained building-boundary details.Source: C. Synergistic Temporal-Spatial… · p. 5
  • CrossAtt adds temporal embeddings, then uses counterpart queries within each feature’s key–value space and a residual connection.The query and key channels are compressed to one-eighth of their original dimensions, while the value dimension is preserved.Source: C. Synergistic Temporal-Spatial… · p. 5
  • The learnable CrossAtt scale ω is initialized to 0 for training stability.Source: C. Synergistic Temporal-Spatial… · p. 5
  • CoordAtt pools features along height and width, fuses them, and uses directional attention maps to reweight the input features.The directional maps are integrated by channel-wise concatenation and convolutional fusion.Source: C. Synergistic Temporal-Spatial… · pp. 5–6

D. Lightweight Gated Fusion Unit (LGFU)

LGFU uses high-level features to control how much shallow spatial detail is fused at each location. Its gated fusion is applied bottom-up through the decoder to form the final high-resolution change map.

D. Lightweight Gated Fusion Unit (LGFU)

  • LGFU gates shallow features pixel-wise, filtering noise and clutter while preserving meaningful change boundaries.The gate controls how much low-level detail flows into the semantic prediction rather than merging all shallow features indiscriminately.Source: D. Lightweight Gated Fusion Unit (LGFU) · p. 6
  • At each decoder step, upsampled deep difference features are aligned with adjacent shallow features and fed together to the gating unit.The unit outputs a single-channel gating map with values in [0, 1].Source: D. Lightweight Gated Fusion Unit (LGFU) · p. 6
  • The gated shallow features are added to the aligned deep features, and fusion proceeds bottom-up until the high-resolution change map is generated.Source: D. Lightweight Gated Fusion Unit (LGFU) · p. 6
  • The network is trained end-to-end with a linear combination of Binary Cross-Entropy and Dice losses to address class imbalance.The passage describes Binary Cross-Entropy as enforcing pixel-level accuracy and Dice loss as robust to skewed distributions.Source: D. Lightweight Gated Fusion Unit (LGFU) · p. 6

A. Datasets

The evaluation uses CDD, GZ-CD, and LEVIR-CD, which cover seasonal and disaster-related changes, season-varying imagery, and building changes. The CDD visualization also encodes prediction outcomes by color.

A. Datasets

  • CDD contains 11 Google Earth image pairs at 0.03–1 m/pixel, with seasonal and disaster-induced changes, and uses 10,000/3,000/3,000 training, validation, and test patches.Its scenarios include shadows, lighting variations, and cloud cover.Source: A. Datasets · p. 6
  • GZ-CD has 1,073 filtered Guangzhou image pairs at 0.55 m/pixel, split into 751/214/108 training, validation, and test pairs.The season-varying images were sourced from Google Earth via BIGEMAP.Source: A. Datasets · p. 6
  • LEVIR-CD comprises 637 building-change image pairs at 0.5 m/pixel, documenting changes over 5–14 years and split into 7,120/1,024/2,048 training, validation, and test patches.The images cover multiple cities across Texas, USA.Source: A. Datasets · p. 6
  • In the CDD result visualization, white/black denote true positives/negatives, red denotes false positives, and green denotes false negatives.Colored dashed boxes highlight differences among algorithms.Source: Figure 4 · p. 7

B. Experimental Setup

The experiments use a single-GPU PyTorch setup with AdamW and cosine learning-rate decay. Evaluation reports five standard metrics, prioritizing F1 and IoU.

1) Implementation Details:

  • AdamW training runs for 100 epochs with cosine decay to 1e-6, using initial learning rates of 1e-3 for the decoder and head and 1e-4 for the pre-trained backbone.Weight decay is 0.01.Source: 1) Implementation Details · p. 7
  • Experiments use PyTorch on one NVIDIA RTX 4090 GPU with batch size 32 across datasets.Source: 1) Implementation Details · p. 7

2) Evaluation Metrics:

  • Evaluation uses Precision, Recall, F1, IoU, and Overall Accuracy, with F1 and IoU prioritized for balanced evaluation.Source: 2) Evaluation Metrics · p. 7

C. Comparisons With SOTAs

FSG-Net outperforms comparison methods on nearly every metric across CDD, GZ-CD, and LEVIR-CD. Qualitative examples highlight pseudo-change suppression and boundary fidelity.

1) Quantitative Evaluation:

2) Qualitative Evaluation:

  • FSG-Net nearly flawlessly traces complex rural-path changes and suppresses illumination-induced pseudo-changes in CDD scenes.The authors describe a few remaining misclassifications in the rural-path example.Source: 2) Qualitative Evaluation · pp. 8–9
  • On GZ-CD, FSG-Net segments new buildings while preserving clear boundaries between changed and unchanged regions.The authors link this example to STSAM’s temporal context modeling and fine-grained edge focus.Source: Figure 5 · p. 8, 2) Qualitative Evaluation · p. 9
  • On LEVIR-CD, FSG-Net better preserves building structures and sharp edges than competing algorithms across varied building forms.Figure 6 visualizes results on LEVIR-CD.Source: 2) Qualitative Evaluation · p. 9, Figure 6 · p. 9

E. Ablation Experiment and Analysis

The ablation analysis evaluates a baseline and individual modules, with DAWIM improving results and combining DAWIM with STSAM producing a greater-than-additive benefit.

E. Ablation Experiment and Analysis

  • The ablation study uses a baseline with basic frequency differencing and SelfAtt in place of DAWIM and STSAM, paired with a U-Net decoder.The stated aim is to distinguish gains from the proposed designs from gains due to generic enhancements.Source: E. Ablation Experiment and Analysis · p. 9

1) Effectiveness of DAWIM:

  • Adding DAWIM to the baseline improves F1 by 1.64% on CDD and 1.83% on GZ-CD.The authors report consistent gains across all three datasets and relate the gains on seasonally variable datasets to reduced task-irrelevant interference.Source: 1) Effectiveness of DAWIM · p. 9
  • DAWIM and STSAM together yield a performance gain greater than the sum of their individual contributions.The authors attribute the cumulative benefit to DAWIM filtering background noise before STSAM focuses attention on genuine change targets.Source: 1) Effectiveness of DAWIM · p. 10
  • DAWIM’s frequency-domain interaction preserves spatial feature patterns while introducing temporal changes in feature intensity and texture.Figure 7 displays normalized features before and after interaction for both temporal images.Source: 1) Effectiveness of DAWIM · p. 10, Figure 7 · p. 10

2) Effectiveness of STSAM:

STSAM combines temporal and spatial attention to refine change features, while LGFU guides shallow features with high-level semantics; ablations and visualizations examine both modules.

2) Effectiveness of STSAM:

  • Removing STSAM from the full model lowers F1 by 1.28% on CDD, 1.56% on GZ-CD, and 0.54% on LEVIR-CD.The module combines CrossAtt with temporal embeddings for global temporal dependencies and CoordAtt for local spatial locations.Source: 2) Effectiveness of STSAM · p. 10
  • STSAM’s temporal and spatial streams refine features to identify multi-scale changes while preserving sharp boundaries, including a building obscured by snow.Figure 8 shows feature refinement in two building-change scenarios.Source: 2) Effectiveness of STSAM · pp. 10–11

3) Effectiveness of LGFU:

  • Adding LGFU to the baseline raises IoU by 1.37% on CDD, 1.26% on GZ-CD, and 0.55% on LEVIR-CD.Removing LGFU from the full model lowers IoU by 1.43% and 1.53% on the first two datasets; the authors report that its contribution is amplified on refined features.Source: 3) Effectiveness of LGFU · p. 11, Table IV · p. 11
  • The complete model produces sharper change boundaries than the baseline and the model without LGFU in the predicted probability heatmaps.The heatmap’s color intensity represents predicted change probability.Source: 3) Effectiveness of LGFU · p. 11

4) Analysis on DAWIM’s Interaction Strategies:

FSG-Net’s evaluations examine frequency-domain interaction strategies, temporal-spatial attention, and model efficiency, alongside its benchmark performance and remaining limitations. The conclusion describes a frequency-spatial approach to suppressing pseudo-changes and a semantics-guided approach to shallow-detail fusion.

4) Analysis on DAWIM’s Interaction Strategies:

  • On CDD, DAWIM exceeds the Difference, Conv2x1x1, and Conv2x3x3 strategies in F1 by 1.33%, 0.69%, and 0.51%, respectively.Removing the residual connection lowers F1 by 1.59%, 1.70%, and 0.33% on the three datasets, respectively.Source: 4) Analysis on DAWIM’s Interaction… · p. 11

5) Comparison on STSAM’s Components:

  • STSAM’s temporal embedding increases F1 by 0.15%, 0.17%, and 0.09% on the three datasets, while combining attention mechanisms leaves F1 lower than CrossAtt alone by 0.22%, 0.19%, and 0.11%.The passage attributes the lower combined-attention performance to functional overlap and describes temporal embedding as beneficial when performance approaches a ceiling.Source: 5) Comparison on STSAM’s Components · p. 11

F. Discussion

  • 94.16% F1 and 88.96% IoU on CDD, 89.51% F1 and 81.01% IoU on GZ-CD, and 91.27% F1 and 83.94% IoU on LEVIR-CD are reported as state-of-the-art results.The discussion also reports failure to detect minute, isolated building changes amid vast homogeneous backgrounds.Source: F. Discussion · p. 12
  • The discussion identifies missed minute, isolated building changes amid vast homogeneous backgrounds as a limitation.It suggests robustness in extreme scenarios could be further enhanced.Source: F. Discussion · p. 12
  • With 13.76M parameters and 6.21 GFLOPs, FSG-Net is reported to balance complexity and efficiency against parameter-heavy CDNeXt and computationally expensive ChangeFormer.Params and FLOPs were measured using a unified 256×256×3 image input size.Source: F. Discussion · p. 12, Table V · p. 12

V. CONCLUSION

  • FSG-Net uses DAWIM to tailor processing to frequency components, STSAM to connect bitemporal features’ attention to genuine changes, and LGFU to guide and filter shallow details with high-level semantics.The conclusion characterizes this as a synergistic frequency-spatial strategy for pseudo-change suppression and semantic-gap handling.Source: §V · p. 12
  • Future work is directed toward lightweight architectures, multimodal data, or knowledge graphs with rule-based constraints for extreme and complex scenarios.Source: §V · p. 12
Loading 2509.06482v1…