Source-linked AI summary
Building Change Detection for Remote Sensing Images Using a Dual Task Constrained Deep Siamese Convolutional Network Model
Yi Liu, Chao Pang, Zongqian Zhan, Xiaomeng Zhang, Xue Yang
TL;DR
Building change detection needs more discriminative features because existing results can contain incomplete regions and irregular boundaries. DTCDSCN jointly performs change detection and semantic segmentation, adds dual attention, and uses improved focal loss; on the WHU building dataset, it reports better performance than other state-of-the-art methods.
Problem
Existing building change detection methods can produce incomplete changed regions and irregular boundaries because extracted features are not discriminative enough.
Method
DTCDSCN combines a change detection network with two semantic segmentation networks, dual attention over channels and spatial positions, and an improved focal-loss-based CDL function.
Results
The proposed method reports better performance than other state-of-the-art methods on the WHU building dataset.
Takeaways & Limitations
Joint semantic segmentation and change detection yields more complete changed areas and more regular boundaries, while CDL improves results under sample imbalance.
Abstract
from arXiv · showhide
In recent years, building change detection methods have made great progress by introducing deep learning, but they still suffer from the problem of the extracted features not being discriminative enough, resulting in incomplete regions and irregular boundaries. To tackle this problem, we propose a dual task constrained deep Siamese convolutional network (DTCDSCN) model, which contains three sub-networks: a change detection network and two semantic segmentation networks. DTCDSCN can accomplish both change detection and semantic segmentation at the same time, which can help to learn more discriminative object-level features and obtain a complete change detection map. Furthermore, we introduce a dual attention module (DAM) to exploit the interdependencies between channels and spatial positions, which improves the feature representation. We also improve the focal loss function to suppress the sample imbalance problem. The experimental results obtained with the WHU building dataset show that the proposed method is effective for building change detection and achieves a state-of-the-art performance in terms of four metrics: precision, recall, F1-score, and intersection over union.
I. INTRODUCTION
Building change detection supports land-use planning, city management, and emergency response, but manual workflows are inefficient and existing methods can produce incomplete, irregular change regions. DTCDSCN addresses these issues by combining change detection with semantic segmentation, dual attention, and an improved focal loss.
- Building change detection identifies changes across remote sensing images acquired at different times and supports land-use planning, city management, and emergency response.
- Manual change detection is time-consuming and labor-intensive, motivating automatic and efficient approaches.
- Existing methods include post-classification comparison and direct classification, but post-classification can accumulate extraction errors and threshold-based features may require many empirical thresholds.
- Deep-learning methods extract more robust features, yet earlier approaches were performed step by step rather than through an ideal unified process.
- Improved SegNet can produce pixel-level building change maps, but may miss small spectral or structural changes and generate holes with irregular boundaries.
- DTCDSCN jointly performs building change detection and building extraction so shared features become more discriminative at the object level.
- The model adds a dual attention module using dependencies between channels and spatial positions, and an improved focal loss to address sample imbalance.
II. PROPOSED METHOD
DTCDSCN unifies change detection and semantic segmentation in a shared framework to focus learning on object-level building features while addressing extreme class imbalance.
- DTCDSCN consists of a change detection network and two semantic segmentation networks operating simultaneously in a unified framework.
- Joint task learning focuses the model on object-level building features and improves feature discriminability for change detection.
- Because changed samples are fewer than unchanged samples, the model develops the CDL function to address extreme sample imbalance.
A. Siamese Change Detection Network
The Siamese change detection network uses shared-weight encoders and an encoder–decoder design, with spatial-channel attention incorporated into decoding to strengthen feature discrimination.
- The change detection network is a fully convolutional Siamese model with two shared-weight encoding branches and one decoding branch.
- Its encoder uses SE-ResNet as the basic module and spatial pyramid pooling to incorporate global context and multiscale features.
- The decoder has five parts and uses the D-LinkNet decoder as its baseline design.
- The decoder inserts DAM into change detection blocks to model correlations across spatial positions and channels.
- DAM combines channel-level and spatial-level attention maps with the input through a shortcut connection to improve feature discriminability.
B. Semantic Segmentation Network
The semantic segmentation network resembles the change detection network but operates on a single temporal image phase, supporting the dual-task design.
- The semantic segmentation network is similar to the change detection network because the two tasks are related.
- Unlike the change detection network, an individual semantic segmentation network applies only to images from a single time phase.
C. Network Architecture
DTCDSCN combines change detection with two semantic segmentation networks in a Siamese framework, while CDL addresses sample imbalance through adjusted changed- and unchanged-sample weighting.
- Network architecture: DTCDSCN integrates a change detection network with two weight-sharing semantic segmentation networks to learn object-level building features.The model receives bitemporal images and outputs a change detection map plus segmentation results for both phases.
- Network architecture: The Siamese change detection network uses two encoders with shared trainable weights for the two image phases and one decoder.The overall network follows a fully convolutional encoder–decoder design.
- Loss function: The loss combines segmentation losses with CDL, weighted by α for segmentation and β for change detection.The cited formulation defines y as ground truth and ŷ as the prediction; α and β weight the two loss components.
- Loss function: CDL improves focal loss by modeling changed- and unchanged-sample weights nonlinearly to address imbalanced sample sets.Its two parameters control the respective sample weights, while focal loss can let easily classified negatives dominate gradients.
III. EXPERIMENTS
The experiments use the WHU building dataset because it provides semantic labels for both dates together with change labels, enabling evaluation of the proposed dual-task method.
- Dataset selection: The WHU building dataset was selected because it provides bitemporal semantic building labels and corresponding change labels.The paper states that other change-detection datasets generally contain only bitemporal images and change labels.
A. Dataset
The WHU dataset contains two Christchurch scenes from 2012 and 2016, tiled into nonoverlapping 256 × 256 patches and divided into training, validation, and test sets.
- Dataset: The dataset covers a Christchurch area affected by a magnitude-6.3 earthquake in February 2011.It contains imagery acquired at the same location in 2012 and 2016, with building semantic labels and change labels.
- Dataset: The 32507 × 15354-pixel images were divided into nonoverlapping 256 × 256-pixel tiles.The resulting sets contain 6096 training, 762 validation, and 762 test tiles.
- Implementation: The models were trained with mini-batch ADAM using batch size 16 and an initial learning rate of 10^-3.PyTorch was used, with two NVIDIA GTX 1080 Ti GPUs.
C. Ablation study
The ablation study on WHU shows incremental gains from DAM, focal loss, and CDL, while adding semantic segmentation improves quantitative and visual change-map quality.
- Study design: The ablation study evaluated each network component on the WHU building dataset using Table I and visualized results in Fig. 3.The paper reports that data augmentation additionally improved F1 by around 1.32% and IoU by 0.8%.
- DAM: DAM increased F1 score from 79.42% to 82.93% and changed-area IoU from 65.86% to 70.84%.DAM exploits dependencies within channels and spatial positions through self-attention.
- Loss functions: Focal loss raised F1 score from 82.93% to 83.48% and IoU from 70.84% to 71.65%.These values are reported in the ablation comparison after adding focal loss to SCDN with DAM.
- Dual-task network: Combining the change detection network with semantic segmentation networks significantly improved all measures.The resulting changed areas were more complete and had more regular boundaries in the visual comparison.
D. Performance Evaluation
The study compares DTCDSCN with other state-of-the-art change detection methods using the WHU building dataset.
- DTCDSCN was evaluated against other state-of-the-art change detection methods on the WHU building dataset.
TABLE I
The evaluation covers change detection, semantic segmentation, and post-classification comparison, while examining label errors and training choices.
- Change detection evaluation: The performance of different change detection methods is reported in Table II after data augmentation.The authors did not use 5-fold cross-validation because training and verification would have required substantially more time.
- Semantic segmentation and post-classification: The proposed method performs semantic segmentation and supports post-classification comparison based on building extraction results.
- Semantic segmentation and post-classification: The proposed method achieves about 5% lower IoU in semantic segmentation than UNet.
- Semantic segmentation and post-classification: Manual label inconsistency and image-registration deviations amplify boundary errors in post-classification comparison, producing 88.86% building-extraction IoU but only 68% change-detection IoU.The WHU change-detection labels are derived from vector objects, which reduces error accumulation relative to subtracting semantic-segmentation labels.
IV. CONCLUSION
The paper proposes DTCDSCN for joint building change detection and semantic segmentation, incorporating attention-based feature representation and imbalance-aware training. On the WHU building dataset, it outperforms other state-of-the-art methods, while future work targets broader and less supervised settings.
- DTCDSCN performs semantic segmentation and change detection within one model, using a dual attention module to improve feature representation.
- The model is trained end-to-end with the CDL function to suppress sample imbalance.
- On the WHU building dataset, DTCDSCN achieves better performance than other state-of-the-art methods.
- Future work will extend the model to general change detection and unsupervised or weakly supervised change detection.