Source-linked AI summary
Change is Everywhere: Single-Temporal Supervised Object Change Detection in Remote Sensing Imagery
Zhuo Zheng, Ailong Ma, Liangpei Zhang, Yanfei Zhong
TL;DR
Pairwise labeling of bitemporal HSR imagery is expensive, motivating learning from unpaired labeled images. STAR constructs supervisory change signals from unpaired images, and ChangeStar implements this approach by reusing semantic segmentation architectures; experiments report strong performance, while false positives remain possible without actual negative samples.
Problem
Deep ConvNet change detectors require large numbers of pairwise labeled bitemporal images, whose annotation is expensive and time-consuming.
Method
STAR exploits object changes in unpaired labeled images, while ChangeStar combines semantic segmentation with ChangeMixin for object change detection.
Results
ChangeStar significantly outperforms PCC across representative segmentation models under single-temporal supervision and shows improvements from semantic supervision and temporal symmetry.
Takeaways & Limitations
STAR provides a competitive object change detection approach using cheaper labels and can reuse modern semantic segmentation architectures.
Takeaways & Limitations
STAR is partly affected by false positives because unpaired training lacks actual negative samples such as the same object at different times.
Abstract
from arXiv · showhide
For high spatial resolution (HSR) remote sensing images, bitemporal supervised learning always dominates change detection using many pairwise labeled bitemporal images. However, it is very expensive and time-consuming to pairwise label large-scale bitemporal HSR remote sensing images. In this paper, we propose single-temporal supervised learning (STAR) for change detection from a new perspective of exploiting object changes in unpaired images as supervisory signals. STAR enables us to train a high-accuracy change detector only using \textbf{unpaired} labeled images and generalize to real-world bitemporal images. To evaluate the effectiveness of STAR, we design a simple yet effective change detector called ChangeStar, which can reuse any deep semantic segmentation architecture by the ChangeMixin module. The comprehensive experimental results show that ChangeStar outperforms the baseline with a large margin under single-temporal supervision and achieves superior performance under bitemporal supervision. Code is available at https://github.com/Z-Zheng/ChangeStar
1. Introduction
Object change detection is valuable for earth-vision applications but conventionally depends on expensive pairwise labeling of coregistered bitemporal images. STAR instead learns from unpaired labeled images by exploiting object changes as supervisory signals, while ChangeStar combines this strategy with reusable segmentation architectures.
- Object change detection outputs pixel-wise changes from bitemporal HSR imagery for applications including urban planning, environmental monitoring, and disaster assessment.
- Deep ConvNet change detectors require many pairwise labeled bitemporal images, making large-scale annotation expensive and time-consuming.
- Bitemporal supervision assigns change labels from semantic differences at corresponding pixels, so positional consistency is needed for accurate supervision.
- STAR uses object changes between unpaired labeled images as supervisory signals, avoiding the need to collect paired labeled images.
- ChangeStar combines an arbitrary deep semantic segmentation model with ChangeMixin, enabling segmentation architectures to detect object changes.
- Temporal symmetry is used as an inductive bias to reduce overfitting caused by the absence of positional consistency in unpaired images.
2. Related Work
Object change detection focuses on whether objects have changed, whereas prior supervised methods generally require change labels from paired images of the same area. Single-temporal segmentation comparison offers a label-efficient baseline but does not model temporal information explicitly.
- Object change detection is object-centric and includes binary change detection, such as building change detection, and semantic change detection, such as damage assessment.
- Supervised object change detection traditionally requires change labels from bitemporal images of the same area, whose pairwise annotation is expensive and time-consuming.
- Deep ConvNet methods commonly use shared encoders and temporal feature differences to detect changes between bitemporal images.
- Post-classification comparison uses single-temporal semantic segmentation outputs as a change-detection baseline but ignores temporal information modeling.
- Bitemporal training samples contain images at times t1 and t2 plus a change label for the intervening period, requiring coregistration for accurate supervision.
3. Approach
The approach replaces paired bitemporal supervision with semantic comparisons between unpaired images, then implements this idea in ChangeStar by combining segmentation with change-specific temporal modeling. STAR constructs pseudo-bitemporal training pairs, assigns change labels from semantic labels, and uses symmetry-aware multi-task supervision.
- Rethinking Bitemporal Supervised Learning: STAR relaxes the same-location requirement by learning from semantic comparisons between two unpaired images.The resulting training formulation can still detect object changes in multi-temporal images of the same area.
- Pseudo-Bitemporal Pair Construction: STAR constructs pseudo-bitemporal pairs by randomly permuting images within a single-temporal mini-batch.The original and permuted sequences replace the two temporal inputs used in conventional training.
- Change Label Assignment: STAR automatically assigns object-change labels from the semantic labels of the paired images, marking pixels where the object appears only once as positive.This avoids manually pairwise labeling dense bitemporal change masks for binary object change.
- Multi-task Supervision: The multi-task objective combines semantic segmentation and change losses to jointly learn object segmentation and object change detection.For binary object segmentation, the semantic loss is binary cross-entropy.
- Temporal Symmetry: Temporal symmetry regularizes binary change detection by requiring predictions to remain consistent when the temporal order of the inputs is reversed.The paper uses this inductive bias to alleviate overfitting.
- ChangeStar Architecture: ChangeStar combines an arbitrary deep semantic segmentation model with ChangeMixin, whose temporal swap module and convolutional layers produce change predictions.The temporal swap module creates two channel-wise temporal permutations, while the small convolutional network uses shared weights during training.
4. Experiments
Experiments evaluate STAR and ChangeStar across cross-domain building-change datasets, component ablations, bitemporal supervision, and learning behavior. ChangeStar consistently benefits from semantic supervision, temporal symmetry, suitable architecture choices, and learned object-change representations.
- Experimental Setting: Experiments use two training segmentation datasets and two evaluation building-change datasets to assess cross-domain object change detection.Models were trained with single-temporal images and semantic labels, then evaluated on WHU building change detection and LEVIR-CD.
- Main Results: ChangeStar significantly outperforms PCC baselines across representative segmentation models under single-temporal supervision.The reported cross-domain evaluation uses only single-temporal supervision, with gains attributed to learning object-change representations rather than only comparing semantic predictions.
- Ablation Study: ChangeMixin performs best with N = 4 convolutional layers and dc = 16 filters, while deeper subnetworks and larger dc values reduce performance.Performance becomes worse than post-classification comparison when N ≥ 6 or dc ≥ 80; N = 4 and dc = 16 are adopted as the default trade-off.
- Ablation Study: Semantic supervision improves the baseline by 0.57% IoU and 0.43% F1, while temporal symmetry adds 2.25% IoU and 1.69% F1 over the baseline.Semantic supervision also improves the temporally symmetric baseline by 1.61% IoU and 1.19% F1; temporal symmetry adds 3.29% IoU and 2.45% F1 over the semantically supervised baseline.
- Bitemporal Supervision: Under bitemporal supervision, encoder-decoder and FPN-based ChangeStars outperform other architectural variants, with deeper backbones providing additional accuracy gains.Semantic FPN and FarSeg are reported as superior among the evaluated variants, while atrous-convolution and spatial-temporal-attention approaches achieve compatible results.
- Error Analysis: STAR remains partly affected by false positives because unpaired supervision lacks actual negative examples such as the same object at different times.Despite this limitation, STAR recognizes many unseen negative examples successfully.
- Learning Behavior: After convergence, ChangeStar’s change prediction surpasses semantic-prediction comparison by a large margin, despite slower early object-change representation learning.Semantic comparison is better during epochs (0, 40], similar during (40, 60], and inferior after convergence.
5. Conclusion
The paper introduces STAR to exploit object changes in unpaired images, avoiding the need to collect pairwise labeled bitemporal images. ChangeStar demonstrates the approach through joint semantic segmentation and object change detection using reusable segmentation architectures.
- 5. Conclusion: STAR bypasses the costly collection of pairwise labeled images by using object changes in arbitrary unpaired images as supervisory signals.The paper positions STAR as a weakly supervised perspective for object change detection.
- 5. Conclusion: ChangeStar is a multi-task architecture for joint semantic segmentation and object change detection that reuses deep semantic segmentation architectures through ChangeMixin.The conclusion describes ChangeStar as simple yet effective and applicable with different segmentation architectures.
- 5. Conclusion: Experiments report competitive performance across different domains with cheaper labels.The paper presents STAR as a baseline intended to support future weakly supervised object change detection research.