PaperlayerView exact revision on arXiv

Source-linked AI summary

Change Detection in Synthetic Aperture Radar Images Using a Dual-Domain Network

Xiaofan Qu, Feng Gao, Junyu Dong, Qian Du, Heng-Chao Li

arXiv:2104.06699v2cs.CVeess.IV

TL;DR

SAR change detection is complicated by speckle noise, while existing methods often emphasize spatial features and patch margins can introduce noise. DDNet combines spatial features, DCT-domain features, and multi-region convolution; experiments report superior performance to several methods on three datasets.

  • Problem

    SAR change detection faces speckle noise, and existing models largely focus on spatial features while patch margins may introduce noisy features.

    Source: §I · p. 1

  • Method

    DDNet combines spatial and DCT-domain features with a multi-region convolution module that emphasizes patch centers while retaining context.

    Source: §I · pp. 1–2

  • Results

    Experiments on three datasets show DDNet is superior to several change-detection methods.

    Source: §I · p. 2, §IV · p. 5

  • Takeaways & Limitations

    The reported experiments support DDNet’s effectiveness for SAR change detection across the evaluated datasets.

    Source: §I · p. 2, §IV · p. 5

    Evaluation is reported on three datasets; the authors identify large-scale dataset validation as future work.

    Source: §IV · p. 5

Abstract

from arXiv · show

Change detection from synthetic aperture radar (SAR) imagery is a critical yet challenging task. Existing methods mainly focus on feature extraction in spatial domain, and little attention has been paid to frequency domain. Furthermore, in patch-wise feature analysis, some noisy features in the marginal region may be introduced. To tackle the above two challenges, we propose a Dual-Domain Network. Specifically, we take features from the discrete cosine transform domain into consideration and the reshaped DCT coefficients are integrated into the proposed model as the frequency domain branch. Feature representations from both frequency and spatial domain are exploited to alleviate the speckle noise. In addition, we further propose a multi-region convolution module, which emphasizes the central region of each patch. The contextual information and central region features are modeled adaptively. The experimental results on three SAR datasets demonstrate the effectiveness of the proposed model. Our codes are available at https://github.com/summitgao/SAR_CD_DDNet.

I. INTRODUCTION

SAR change detection is challenged by speckle noise, while conventional difference-image analysis and many deep-learning methods have limitations in exploiting rich features across domains. DDNet combines spatial and DCT-domain features with a multi-region convolution module, and experiments on three real SAR datasets demonstrate its effectiveness.

I. INTRODUCTION

  • Existing deep-learning models largely focus on spatial features, while patch-wise analysis can introduce noisy marginal features and SAR speckle complicates change detection.Traditional difference-image methods capture some pixel-wise changes but do not adaptively exploit rich representations in the original data.Source: §I · p. 1
  • DDNet integrates reshaped DCT coefficients as a frequency branch alongside spatial features to address the limited use of frequency information in SAR change detection.The authors identify this as the first use of DCT-domain features for SAR change detection and state that the dual-domain representations alleviate speckle noise.Source: §I · pp. 1–2
  • The multi-region convolution module emphasizes each patch’s central area while retaining contextual information, adaptively modeling both for classification.This design targets noisy features that may enter from patch margins.Source: §I · pp. 1–2
  • Experiments on three real SAR datasets demonstrate the effectiveness of DDNet, and the authors made its PyTorch code publicly available.The introduction reports promising results but does not specify quantitative performance values here.Source: §I · p. 2

II. METHODOLOGY

The task is to produce a binary change map from two SAR images of the same region captured at different times. DDNet uses spatial and frequency-domain branches, and its training and classification procedure operates without labeled samples.

II. METHODOLOGY

  • The unsupervised workflow uses log-ratio differences and hierarchical FCM clustering to identify reliable changed and unchanged training pixels and uncertain pixels for later classification.Balanced training samples are randomly drawn from 10% of the image patches centered on reliable pixels; the network then classifies patches centered on uncertain pixels.Source: §II · p. 2
  • For each pixel, patches from both time points are combined into a 2 × r × r input, and the resulting change map marks changed pixels as 1 and unchanged pixels as 0.The patch size is r × r, with r = 7 in this work.Source: §II · p. 2
  • DDNet has a spatial branch for multi-region features and a frequency branch for DCT coefficients.The spatial branch contains four MRC modules; the frequency branch applies DCT and an on-off switch to select critical coefficients.Source: Figure 1 · p. 2, §II · p. 2

A. Spatial Domain Feature Extraction

The spatial branch uses multi-region convolution to emphasize patch centers while retaining context. It creates global and center-focused feature groups, processes them with convolution, and fuses their outputs.

A. Spatial Domain Feature Extraction

  • The three groups are processed by a 3 × 3 convolution and their resulting features are fused by element-wise summation.The MRC illustration shows a patch-size example of r = 7 with 15 input channels.Source: Figure 2 · p. 2, A. Spatial Domain Feature Extraction · p. 3
  • Three regional views—global, horizontal middle, and vertical middle—retain context while emphasizing the patch center and suppressing noisy marginal pixels.The horizontal view removes top and bottom rows; the vertical view removes left and right columns.Source: A. Spatial Domain Feature Extraction · p. 2
  • A 1 × 1 convolution maps each input patch to feature maps that are split into global, horizontal-middle, and vertical-middle channel groups.The horizontal group zeros several top and bottom rows, while the vertical group zeros several left and right columns; for r = 7, two rows or columns are removed on each side.Source: A. Spatial Domain Feature Extraction · pp. 2–3
  • The fused spatial features are reshaped into a 245-element vector with global context and enhanced center-region information.With C = 15, the fused feature map is 5 × 7 × 7.Source: A. Spatial Domain Feature Extraction · p. 3

B. Frequency Domain Feature Extraction

DCT is used to extract frequency features because compressed frequency representations can suppress speckle noise. The method transforms resized image patches and gates informative coefficients to form a frequency feature vector.

B. Frequency Domain Feature Extraction

C. Final Change Map Generation

The spatial and frequency features are combined to classify pixels and generate the final change map.

C. Final Change Map Generation

  • Spatial and frequency features are concatenated, passed through a fully connected layer and softmax classifier, and used to classify pixels into the final change map.The softmax estimates changed versus unchanged probabilities.Source: C. Final Change Map Generation · p. 3

III. EXPERIMENTAL RESULTS AND ANALYSIS

The experiments use three real SAR datasets to assess DDNet, examine parameters affecting change detection, and compare it with state-of-the-art methods.

III. EXPERIMENTAL RESULTS AND ANALYSIS

  • Experiments on three real SAR datasets assess evaluation criteria, discuss parameters affecting change detection, and compare DDNet with state-of-the-art methods.The section describes datasets and criteria before parameter analysis and method comparisons.Source: §III · p. 3

A. Dataset and Evaluation Metric

The evaluation uses three co-registered, geometrically corrected multitemporal SAR datasets from different sensors and five common change-detection metrics.

A. Dataset and Evaluation Metric

  • The study uses co-registered and geometrically corrected multitemporal datasets from Ottawa, Sulzberger, and the Yellow River, acquired by different sensors.The datasets are described as real SAR imagery.Source: A. Dataset and Evaluation Metric · p. 3
  • Ottawa is 290 × 350 RADARSAT imagery from May and August 1997; Sulzberger is 256 × 256 Envisat imagery of the ice-shelf breakup caused by a March 2011 tsunami.Sulzberger is a part of the Sulzberger Ice Shelf.Source: A. Dataset and Evaluation Metric · p. 3
  • The 291 × 306 Yellow River dataset contains Radarsat-2 imagery from June 2008 and June 2009, where stronger speckle makes changed regions difficult to identify accurately.Source: A. Dataset and Evaluation Metric · p. 3
  • Performance is measured using false positives (FP), false negatives (FN), overall error (OE), percentage of correct classification (PCC), and kappa coefficient (KC).Source: A. Dataset and Evaluation Metric · p. 3

B. Analysis of the Patch Size

Patch size affects the balance between spatial context, PCC, computation, and noise. The ablation visualization indicates that dual-domain features form clearer, more separated clusters than the input features.

B. Analysis of the Patch Size

  • As patch size r increases, PCC first rises and then stabilizes, indicating that spatial context matters for change detection.The study evaluates r = 5, 7, 9, 11, 13, and 15.Source: B. Analysis of the Patch Size · p. 3
  • The study selects r = 7 because larger patches increase computation and may introduce noise that affects detection performance.The selected size is used in subsequent implementation.Source: B. Analysis of the Patch Size · p. 3

C. Ablation Study

  • On Ottawa, dual-domain feature representations are more separated and clearly bounded than the input features.Fig. 5 presents input, spatial-domain, frequency-domain, and dual-domain representations.Source: C. Ablation Study · p. 4, Figure 5 · p. 4

D. Performance Comparison

DDNet achieves the best performance across the three datasets, surpassing the compared baselines. The reported KC gains vary by dataset, with the largest stated gains on Yellow River.

D. Performance Comparison

  • On Ottawa, DDNet's KC improves by 3.04%, 0.39%, 0.01%, 0.23%, and 0.3% over PCAKM, NR-ELM, DBN, DCNet, and MSAPNet, respectively.The passage also reports that PCAKM misses many changed regions there.Source: D. Performance Comparison · p. 4
  • On Sulzberger, DDNet's KC increases by 0.99%, 0.53%, and 0.57% over DBN, DCNet, and MSAPNet, respectively.The passage notes noisy areas in PCAKM and NR-ELM maps, and missed changed regions in NR-ELM, DBN, and DCNet maps.Source: D. Performance Comparison · pp. 4–5
  • On Yellow River, DDNet gains 8.63%, 10.18%, 3.04%, 0.79%, and 1.01% in KC over PCAKM, NR-ELM, DBN, DCNet, and MSAPNet, respectively.The passage describes Yellow River as having much stronger speckle noise, making changed regions harder to identify accurately.Source: D. Performance Comparison · p. 5
  • DDNet achieves the best performance and substantially surpasses all baselines across the three datasets.Table II presents change-detection results for different methods on the three datasets.Source: D. Performance Comparison · p. 5, Table II · p. 5

IV. CONCLUSION AND FUTUER WORK

DDNet fuses spatial- and frequency-domain features for SAR change detection, using MRC to enhance central-region features and DCT with gating to extract frequency features. Experiments on three datasets show superiority to several change-detection methods; large-scale dataset evaluation is planned.

IV. CONCLUSION AND FUTUER WORK

  • DDNet fuses spatial- and frequency-domain features to improve classification performance for SAR change detection.The spatial branch uses a multi-region convolution module to enhance central-region features; the frequency branch uses DCT and a gating mechanism.Source: §IV · p. 5
  • Experiments on three datasets show DDNet is superior to several change-detection methods.Source: §IV · p. 5
  • The authors plan to verify DDNet on a large-scale dataset in future work.Source: §IV · p. 5

Related papers

Loading 2104.06699v2…