PaperlayerView exact revision on arXiv

Source-linked AI summary

Fully Convolutional Change Detection Framework with Generative Adversarial Network for Unsupervised, Weakly Supervised and Regional Supervised Change Detection

Chen Wu, Bo Du, Liangpei Zhang

arXiv:2201.06030v1cs.CVcs.AIeess.IV

TL;DR

The paper addresses the dependence of unsupervised change detection on pre-detection methods and the difficulty of obtaining abundant fully supervised labels. It proposes FCD-GAN, a unified framework using a segmentor, generator, and discriminator for multiple supervision settings. Experiments indicate effectiveness in unsupervised, weakly supervised, and regional supervised change detection, while the authors note that the basic modules and parameter settings were not explored in detail.

  • Problem

    End-to-end change-detection networks are mainly supervised, while unsupervised methods depend on pre-detection results and abundant fully supervised labels are difficult to obtain.

  • Method

    FCD-GAN combines a segmentor, image-to-image generator, and discriminator, using iterative or adversarial optimization according to the supervision setting.

  • Results

    Experiments indicate FCD-GAN is effective for unsupervised, weakly supervised, and regional supervised change detection, with improvement in USCD and comparable performance to CAM in WSCD.

  • Takeaways & Limitations

    The framework unifies unsupervised, weakly supervised, regional supervised, and fully supervised change detection and introduces definitions for the first three tasks.

  • Takeaways & Limitations

    The paper does not explore detailed network modules or parameter settings, using basic segmentor, generator, and discriminator structures.

Abstract

from arXiv · show

Deep learning for change detection is one of the current hot topics in the field of remote sensing. However, most end-to-end networks are proposed for supervised change detection, and unsupervised change detection models depend on traditional pre-detection methods. Therefore, we proposed a fully convolutional change detection framework with generative adversarial network, to conclude unsupervised, weakly supervised, regional supervised, and fully supervised change detection tasks into one framework. A basic Unet segmentor is used to obtain change detection map, an image-to-image generator is implemented to model the spectral and spatial variation between multi-temporal images, and a discriminator for changed and unchanged is proposed for modeling the semantic changes in weakly and regional supervised change detection task. The iterative optimization of segmentor and generator can build an end-to-end network for unsupervised change detection, the adversarial process between segmentor and discriminator can provide the solutions for weakly and regional supervised change detection, the segmentor itself can be trained for fully supervised task. The experiments indicate the effectiveness of the propsed framework in unsupervised, weakly supervised and regional supervised change detection. This paper provides theorical definitions for unsupervised, weakly supervised and regional supervised change detection tasks, and shows great potentials in exploring end-to-end network for remote sensing change detection.

1 INTRODUCTION

The paper presents FCD-GAN as a unified framework for unsupervised, weakly supervised, regional supervised, and fully supervised change detection. It addresses limits of pre-detection-dependent unsupervised methods and the difficulty of obtaining abundant pixel-level labels by defining task-specific constraints and using segmentor, generator, and discriminator components.

  • Motivation: Existing unsupervised deep-learning methods depend heavily on pre-detection results and do not provide an advanced end-to-end model.These methods use pre-detected change and non-change patches as training samples for feature extraction and semantic discrimination.
  • Motivation: Fully supervised change detection uses end-to-end segmentation networks but requires abundant pixel-level labels that are laborious to produce.Change detection requires interpreting multi-temporal image pairs, making comprehensive labeling difficult in practical applications.
  • Task definitions: Unsupervised change detection constrains the model to find a segmentation region whose removal makes transitions between multi-temporal images more predictable.The formulation relies on the prior that unchanged landscapes can be predicted across images, whereas changed landscapes cannot.
  • Task definitions: Weakly supervised change detection uses only image-pair change or non-change labels and seeks exact changed regions through a constraint based on masked pairs being classified as unchanged.The supplied definition distinguishes image-level supervision from the desired pixel-level change map.
  • Task definitions: Regional supervised change detection provides a probable change region, such as a rectangle, and trains the model to identify the exact changed pixels within it.The task is defined so that masking the predicted region makes the pair appear unchanged, like masking a predefined reference region.
  • Proposed framework: FCD-GAN combines a basic segmentor, an image-to-image generator, and a discriminator to support four change-detection supervision settings in one framework.The segmentor obtains change maps; iterative segmentor-generator optimization supports unsupervised detection, while adversarial constraints support weakly and regionally supervised tasks.

2 METHODOLOGY

FCD-GAN combines a segmentor, generator, and discriminator to support unsupervised, weakly supervised, and regional supervised change detection. Different optimization constraints and adversarial processes adapt the common framework to each task.

  • Framework: The framework contains three basic modules: a segmentor, generator, and discriminator, using simple network structures because the paper focuses on a common framework.The segmentor produces change maps; the generator models image transitions; and the discriminator distinguishes changed from unchanged pairs.
  • Implementation: The implementation uses a basic Siamese U-Net segmentor, an SRGAN-like generator with reconstruction and VGG16 content losses, and iterative Adam optimization.The segmentor concatenates multi-temporal features, while the generator omits a final sigmoid or tanh because inputs are standard-normalized.
  • Unsupervised Change Detection: The unsupervised objective requires sparse segmentation because labeling every pixel as changed would satisfy the prediction objective but produce a false solution.The relaxed loss weight controls the precision–recall trade-off: higher weight generally yields fewer predicted changes, higher precision, and lower recall.
  • Unsupervised Change Detection: Unsupervised change detection iteratively optimizes the segmentor and generator so unchanged regions support accurate prediction between multi-temporal images.The generator predicts one image from another, while the segmentor identifies a change region whose removal improves prediction.
  • Weakly Supervised Change Detection: Weakly supervised change detection uses image-pair change labels and adversarial training between a segmentor seeking changed regions and a discriminator distinguishing masked changed from unchanged pairs.Applying the change mask to unchanged pairs prevents the discriminator from learning the presence of a mask instead of semantic change.
  • Regional Supervised Change Detection: Regional supervised change detection extracts changed pixels inside a probable labeled region through an adversarial process using a simulated unchanged pair masked by the reference region.The model is intended to find sparse changes inside the reference region and no changes outside it.

3 DATASETS

The experiments use GF-2 image pairs for unsupervised detection and construct weakly and regionally supervised datasets from existing change-detection resources. The datasets span optical imagery with pixel-, pair-, and region-level supervision.

  • Unsupervised Dataset: The unsupervised experiments use two high-resolution multi-temporal GF-2 datasets acquired on April 4 and September 1, 2016.The imagery has four bands—Blue, Green, Red, and NIR—and 4m spatial resolution.
  • Dataset Visualization: Figures 5 and 6 show the GF-2 time-1 image, time-2 image, and reference, where red marks changed pixels and green marks unchanged pixels.The figures correspond to the Wuhan and Hanyang multi-temporal datasets.
  • Supervised Dataset Construction: Weakly and regionally supervised datasets are generated from shared supervised change-detection datasets because corresponding datasets are unavailable.The weakly supervised dataset comes from WHU BCD, while OSCD is used for regional supervised detection.
  • Weakly Supervised Dataset: The WHU building change detection dataset contains 2012 and 2016 aerial images with three optical bands at 0.3m resolution.Images are sliced into 200 × 200 patches, producing 9,935 unchanged pairs and 2,616 changed pairs for weak supervision.
  • Dataset Visualization: Figure 7 presents changed and unchanged image-pair samples used for weakly supervised change detection.The WHU-derived dataset is described as complex because some landscape changes are not target semantic changes.
  • Dataset Visualization: Figure 8 shows a Rio image pair acquired on April 24, 2016 and October 11, 2017, alongside pixel-level and region-level references.The region-level reference supports the regional supervised task.
  • Regional Supervised Dataset: The OSCD dataset contains 24 Sentinel-2 multi-temporal pairs, with 14 training and 10 test pairs, for regional supervised evaluation.The experiments use four bands—B, G, R, and NIR—at 10m resolution, with co-registration applied to reduce mis-registration errors in some pairs.

4 EXPERIMENTS

Experiments evaluate FCD-GAN across unsupervised, weakly supervised, and regional supervised change detection. Results show effective change localization, competitive weakly supervised accuracy, and strong regional supervised performance under limited supervision, while also revealing semantic false alarms and remaining WSCD limitations.

  • Experimental setup: Overlapping patches enable fully convolutional processing of large images, using only centered outputs to reduce patch-edge effects.USCD experiments use 220 × 220 patches with 10-pixel overlap and retain the centered 200 × 200 region.
  • Unsupervised change detection: FCD-GAN converges during iterative optimization, and performs best across most thresholds around 0.5 on WH and HY datasets.The generation-loss curves converge after iteration, while threshold comparisons are made against SiamCRNN_LSTM.
  • Unsupervised change detection: FCD-GAN finds most changes and highlights real changes clearly in WH and HY visual assessments.Figures 9 and 10 visualize true detections, false alarms, omission errors, and change-probability outputs.
  • Weakly supervised change detection: FCD-GAN with adversarial processing matches or exceeds CAM on BCD, while adding a small generation-loss weight yields the highest Kappa, F1, mIOU, and cIOU.Generation loss alone has low precision because it detects landscape changes beyond the target semantic changes.
  • Weakly supervised change detection: The adversarial process remains balanced because the discriminator struggles to distinguish real unchanged pairs from segmentor-produced masked pairs.The segmentor loss c converges around 0.5 and discriminator loss d_u around 1.
  • Weakly supervised change detection: FCD-GAN assigns more precise change outlines than CAM but can highlight non-target changed vehicles and other semantic false alarms.CAM produces more integrated regional results, whereas FCD-GAN covers more of the changed-landscape shape with high density.
  • Regional supervised change detection: On OSCD, FCD-GAN achieves the highest accuracy among the evaluated regional supervised variants, with an F1-score of 0.3508 despite limited supervision.The FSCD model reports an accuracy of 0.4044, while FCD-GN and FCD-AN also show satisfactory performance.
  • Regional supervised change detection: A threshold of 0.3 gives the best performance in the regional supervised experiment, while FCD-GAN tends to over-highlight changes relative to ground truth.Region-level supervision identifies some non-changes outside the reference region but cannot exclude every non-target change type.

5 DISCUSSIONS

The discussion distinguishes change-detection supervision by the constraints imposed on a shared segmentor. It argues that weaker supervision is easier to label but generally requires more samples and remains constrained by basic network design choices.

  • Supervision trade-offs: RSCD and WSCD reduce annotation burden because interpreters can provide regional or pair-level supervision more quickly than exhaustive pixel labels.The paper presents these tasks as practical alternatives for collecting many training samples.
  • Supervision trade-offs: The theoretical distinction from FSCD is a shift from labeling target changes to labeling semantic non-changes.FSCD labels changed pixels directly, whereas weaker tasks constrain which changes should be treated as unchanged.
  • Supervision trade-offs: With five landscape classes, 25 possible from-to transitions include only five unchanged types, leaving weaker supervision to represent many non-change cases.RSCD and WSCD must provide remaining transition examples in unchanged pairs or outside reference regions.
  • Supervision trade-offs: RSCD and WSCD may need abundant samples to approach FSCD performance, even though their individual samples are easier to label.Their lower accuracy is attributed to receiving less supervision than FSCD.
  • Unified framework: FCD-GAN separates the segmentor from task-specific constraints: the generator models temporal relationships for USCD, while adversarial discrimination injects semantic supervision for WSCD and RSCD.Fully supervised detection instead uses pixel-level losses such as BCE, DICE, or focal loss.
  • Scope and future development: The framework is demonstrated with basic segmentor, generator, and discriminator modules rather than detailed architectural or parameter exploration.The authors leave advanced fusion, attention, transformer, deep-supervision, and GAN variants for future performance improvements.
  • Contribution: The paper’s main contribution is a general segmentor-generator-discriminator framework covering unsupervised, weakly supervised, regional supervised, and fully supervised change detection.It also defines WSCD and RSCD and details their experimental procedures.

6 CONCLUSION

The paper presents FCD-GAN as an end-to-end framework spanning unsupervised, weakly supervised, regional supervised, and fully supervised change detection, with reported effectiveness in the first three tasks. It also identifies incomplete changed-object regions in WSCD and limitations in the evaluation data as open problems.

  • Conclusion: FCD-GAN addresses unsupervised, weakly supervised, regional supervised, and fully supervised change detection within one framework.WSCD and RSCD are newly defined in the paper, while FSCD is included but not the focus.
  • Conclusion: Iterative segmentor–generator optimization supports USCD, adversarial segmentor–discriminator training supports WSCD and RSCD, and pixel-level segmentor loss supports FSCD.
  • Conclusion: Experiments indicate effectiveness in USCD, WSCD, and RSCD; USCD improves over previous methods, WSCD is comparable to classical CAM, and RSCD achieves satisfactory performance.The paper describes the WSCD result as the first attempt to use GAN in one-stage weakly supervised segmentation and RSCD as a new exploration.
  • Limitations: WSCD results have precise outlines but still contain independent pixels inside segmented regions, leaving changed-object completeness unresolved.
  • Limitations: WSCD and RSCD experiments use data generated from shared FSCD datasets, whose accuracies are reported as not very high.The paper calls for datasets better suited to evaluating WSCD and RSCD algorithms.
  • Conclusion: The paper identifies GAN use in weakly supervised segmentation as a basis for testing the structure on other computer-vision datasets.
Loading 2201.06030v1…