PaperlayerView exact revision on arXiv

Source-linked AI summary

From W-Net to CDGAN: Bi-temporal Change Detection via Deep Learning Techniques

Bin Hou, Qingjie Liu, Heng Wang, Yunhong Wang

arXiv:2003.06583v1cs.CVeess.IV

TL;DR

Remote-sensing change detection is limited by image-domain differencing and hand-crafted features. The paper proposes W-Net, a dual-branch feature-learning network, and CDGAN, which uses W-Net as a GAN generator, and reports effective results with CDGAN improving over W-Net.

  • Problem

    Traditional change detection relies on image differencing and hand-crafted features, limiting representation of complex change information.

  • Method

    The paper proposes dual-branch W-Net for feature-domain change detection and CDGAN, which uses W-Net as a GAN generator for image translation.

  • Results

    Experiments on Google Earth and GF-2 show effective results for W-Net and CDGAN, with CDGAN improving over W-Net.

  • Takeaways & Limitations

    The proposed networks generate change maps directly from the two original images and provide fine-grained change-detection results.

Abstract

from arXiv · show

Traditional change detection methods usually follow the image differencing, change feature extraction and classification framework, and their performance is limited by such simple image domain differencing and also the hand-crafted features. Recently, the success of deep convolutional neural networks (CNNs) has widely spread across the whole field of computer vision for their powerful representation abilities. In this paper, we therefore address the remote sensing image change detection problem with deep learning techniques. We firstly propose an end-to-end dual-branch architecture, termed as the W-Net, with each branch taking as input one of the two bi-temporal images as in the traditional change detection models. In this way, CNN features with more powerful representative abilities can be obtained to boost the final detection performance. Also, W-Net performs differencing in the feature domain rather than in the traditional image domain, which greatly alleviates loss of useful information for determining the changes. Furthermore, by reformulating change detection as an image translation problem, we apply the recently popular Generative Adversarial Network (GAN) in which our W-Net serves as the Generator, leading to a new GAN architecture for change detection which we call CDGAN. To train our networks and also facilitate future research, we construct a large scale dataset by collecting images from Google Earth and provide carefully manually annotated ground truths. Experiments show that our proposed methods can provide fine-grained change detection results superior to the existing state-of-the-art baselines.

I. INTRODUCTION

The paper addresses remote-sensing change detection limitations caused by image-domain differencing and hand-crafted features. It proposes W-Net and CDGAN, and introduces a large annotated dataset for training and evaluation.

  • Motivation: Existing methods commonly preprocess images, compute pixel- or segment-wise difference images, extract change features, and classify the resulting maps.The paper identifies hand-crafted representations and simple image-domain differencing as limitations for complex, high-level change information.
  • W-Net: W-Net independently processes the two bi-temporal images in a dual-branch network, learning feature differences and directly generating change maps.The approach uses a fully convolutional encoder-decoder with feature-domain comparison and feature fusion.
  • CDGAN: CDGAN uses W-Net as a GAN generator, reformulating change detection as image translation and learning a distribution connection between inputs and difference images.The authors motivate adversarial learning as a way to enhance feature representation and generalization.
  • Dataset: 29,000 annotated 256×256 image-patch pairs from 29 Google Earth image pairs form the proposed large-scale change-detection dataset.The patches were carefully annotated by experts to support supervised and deep-learning-based change detection.

II. RELATED WORK

Prior remote-sensing change detection uses pixel- or object-based comparisons, while newer work applies deep networks and GANs. These approaches motivate learning richer representations and mappings for heterogeneous or complex imagery.

  • Traditional methods: Traditional methods compare pixels or segmented objects using differencing, ratioing, regression, CVA, or contextual object analysis.Pixel-based methods may omit local information, whereas object-based methods analyze segmented meaningful regions.
  • Deep learning: Deep-learning approaches apply RBMs, PCANet, multi-resolution feature learning, DBNs, 2D CNNs, and siamese networks to remote-sensing change detection.These methods use learned representations or feature-change analysis across several image settings.
  • GANs: GAN-based work models difference-image distributions or translates between heterogeneous modalities, but cited architectures include simple fully connected layers.The paper positions its conditional adversarial design within this emerging line of research.

III. PROPOSED METHODS

The proposed formulation treats change detection as dense prediction from two images acquired at different times over the same site. The model estimates per-pixel change probabilities and learns parameters by back propagation.

  • 1) Formulation:: Given two bi-temporal images and a ground-truth change map, the task identifies changed areas at corresponding image locations.Binary change values assign 1 to changed pixels and 0 to unchanged pixels.
  • 1) Formulation:: The model replaces binary pixel classification with dense prediction of each pixel’s probability of being changed.This formulation produces a change map from the two input images rather than a single image-level label.
  • 1) Formulation:: The change-detection objective is optimized over model parameters using the stated loss and back propagation.The loss compares target label vectors with predicted label vectors across training samples.

2) Architecture:

W-Net uses a dual-branch encoder-decoder architecture for processing the two images while preserving information through strided convolutions, deconvolutions, concatenation, and shortcuts.

  • 2) Architecture:: A dual-branch design keeps the two input images separate during feature extraction to avoid information loss from early input merging.The branches later support learned comparison and change-map generation.
  • 2) Architecture:: Strided convolution performs downsampling, while deconvolutional layers perform upsampling in the encoder-decoder pathway.The architecture includes layers with and without spatial resolution changes.
  • 2) Architecture:: The architecture uses four convolutional encoder blocks and four deconvolutional decoder blocks, with three shortcut connections between them.Each encoder and decoder block combines stride-1 and stride-2 operations.
  • 2) Architecture:: The W-Net parameter specification is presented in Table I, while its training loss compares target and predicted label vectors over N samples.The supplied table passage identifies the parameter summary but does not provide individual parameter values.

B. GAN Based Change Detection (CDGAN)

CDGAN reformulates change detection as conditional image translation, using W-Net as its generator and a convolutional discriminator to distinguish generated from ground-truth change maps.

  • B. GAN Based Change Detection (CDGAN): CDGAN learns change maps from two bi-temporal images and noise through a conditional mapping G: (X_t1, X_t2, z) → CM.The formulation treats change-map prediction as conditional density estimation.
  • B. GAN Based Change Detection (CDGAN): CDGAN trains generator and discriminator through a minimax objective in which the discriminator distinguishes real data from generator outputs.The generator produces samples from an observed image and random noise, whereas the discriminator separates real and generated samples.
  • B. GAN Based Change Detection (CDGAN): The proposed W-Net serves as CDGAN’s generator, while the discriminator receives both input images and a generated or ground-truth change map.The discriminator outputs 1 or 0 for changed or unchanged classes.

IV. EXPERIMENTS

The experiments introduce the data, implementation and evaluation procedures used to compare W-Net and CDGAN with established baselines.

  • IV. EXPERIMENTS: The experimental study compares W-Net and CDGAN with well-established baselines to validate their effectiveness.The section covers data collection, implementation details and evaluation metrics before the comparison.
  • IV. EXPERIMENTS: The authors construct a large-scale change-detection dataset with high intra-class diversity, low inter-class dissimilarity and ground-truth annotations, and test on Google Earth and GF-2 datasets.The stated purpose is to facilitate training and benefit future research.

1) Data Collection:

The study builds a manually annotated Google Earth training resource from bi-temporal high-resolution imagery and evaluates on separate Google Earth and GF-2 image-pair datasets.

  • 1) Data Collection:: The training resource contains 29,000 pairs of 256×256 patches derived from 29 bi-temporal image pairs collected over Beijing and Tianjin.The source images span 2006–2017 at 0.46 m spatial resolution and include varied acquisition conditions.
  • 1) Data Collection:: The 29,000 patch pairs are split into 23,200 training pairs and 5,800 validation pairs.The split uses 80% for training and 20% for validation.
  • 1) Data Collection:: Testing uses 13 Google Earth image pairs and six GF-2 image pairs, each sized 500×500 pixels and covering Beijing.The Google Earth pairs were captured from 2009 to 2015; the GF-2 data include imagery captured in 2005.
  • 1) Data Collection:: Five expert annotators manually labeled the 29 large image pairs, and three specialists checked the annotations for accuracy.The labeling took about one month and the inspection about two weeks.
  • 1) Data Collection:: Annotators overlaid each bi-temporal pair in ENVI and selected change areas manually with region-of-interest polygons designed to respect ground-object boundaries.Testing data were annotated using similar procedures.

B. Implementation

The implementation uses TensorFlow and Adam, while evaluation combines error-rate, kappa and FAR–MAR metrics to assess detection quality.

  • B. Implementation: The models are implemented in TensorFlow and trained on a GeForce GTX 1080Ti GPU with Adam, an initial learning rate of 0.0002 and adaptive reduction.Weights use zero-mean Gaussian initialization with standard deviation 0.02, biases are initialized to zero, and momentum is 0.5.
  • C. Evaluation Metrics: Evaluation reports missed alarm rate, false alarm rate, overall error rate and kappa coefficient as quantitative metrics.Higher kappa indicates better classification performance.
  • B. Implementation: Table II reports model sizes and training times for W-Net and CDGAN.Its time notation distinguishes days, hours and minutes.
  • C. Evaluation Metrics: MAR, FAR and OER quantify false negatives, false positives and total classification errors using TP, TN, FP and FN.The formulas are MAR = FN/(TP + FN), FAR = FP/(FP + TN), and OER = (FP + FN)/(TP + TN + FP + FN).
  • C. Evaluation Metrics: FAR–MAR curves summarize global detection performance, with smaller AUC generally indicating a higher-quality change map.Better curves lie closer to the coordinate system’s bottom-left corner, where the horizontal axis is MAR and the vertical axis is FAR.

D. Comparison of W-Net and CDGAN

CDGAN generally outperforms W-Net across both testing datasets, although W-Net has a smaller FM-curve AUC and CDGAN has slightly worse MAR. Qualitative results attribute CDGAN’s advantage mainly to fewer false detections and clearer changed regions.

  • CDGAN performs better than W-Net on both testing datasets for all reported metrics except MAR, where its difference is almost negligible.
  • CDGAN’s PR curve is generally above W-Net’s and closer to the top-right corner, whereas W-Net has the smaller FM-curve AUC.
  • Both methods perform better on Google Earth than GF-2, possibly because GF-2 has more complex spectral distributions and strided convolution may overfit Google Earth.
  • CDGAN produces fewer false detections and clearer results than W-Net, while both methods miss little changed area relative to ground truth.
  • The authors report that CDGAN generally achieves better experimental results, attributing this to GAN-based distribution learning and improved transferability and generalization.

E. Comparison of Proposed Methods with Other Baselines

Against seven established baselines, CDGAN achieves the strongest reported OER and kappa performance on both testing datasets, while the proposed methods produce clearer change maps. Their MAR performance remains less competitive, but the authors report overall gains over state-of-the-art methods.

  • CDGAN outperforms all compared methods in OER and kappa on both testing datasets, with W-Net ranking second best.
  • CDGAN ranks first in FAR on Google Earth and second on GF-2, while W-Net ranks third on both datasets.
  • The proposed methods do not achieve competitive MAR performance, possibly because of limited training data, no data augmentation, and randomly initialized non-pretrained backbones.
  • W-Net and CDGAN produce clearer unchanged regions and more accurate changed-region boundaries than the seven baseline methods in qualitative comparisons.
  • The authors conclude that W-Net and CDGAN outperform state-of-the-art baselines and generalize similarly well to the heterogeneous GF-2 dataset.

V. CONCLUSION

The paper proposes W-Net and CDGAN for high-resolution remote sensing change detection, using deep architectures to extract and classify change features. Experiments on two datasets show effective results, with CDGAN improving performance over W-Net and producing change maps directly from original images.

  • W-Net uses a dual-branch architecture with strided convolution, concatenation, and shortcut connections for change-feature extraction and classification.The two branches process bi-temporal images, while fused features support direct change-map generation.
  • CDGAN incorporates W-Net as its generator to learn a mapping function for change detection.
  • Experiments on Google Earth and GF-2 datasets show effective results for both methods, with CDGAN improving performance over W-Net.The evaluation includes qualitative and quantitative comparisons.
  • Unlike traditional methods, the proposed networks generate final change maps directly from the two original images.
Loading 2003.06583v1…