Source-linked AI summary
Fully Convolutional Siamese Networks for Change Detection
Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch
TL;DR
Change detection needs efficient pixelwise analysis of large Earth-observation image collections despite limited annotated datasets. The paper proposes three fully convolutional, end-to-end architectures—including two Siamese variants—and reports improved accuracy and inference speed over prior methods. The models learn from annotated change-detection data without pretraining, while their evaluation scope leaves larger datasets and additional modalities or change types for future testing.
Problem
Large Earth-observation collections require efficient change detection, but limited annotated datasets constrain model development and motivate alternatives to transfer learning.
Method
The paper proposes three fully convolutional encoder-decoder architectures trained end-to-end from scratch, including Siamese networks with shared-weight streams and skip connections.
Results
The architectures surpassed previous change-detection methods in accuracy and inference speed; inference was below 0.1 s and over 500x faster than a cited SCCN setup.
Takeaways & Limitations
Fully convolutional Siamese designs provide a fast change-detection approach that learns directly from available change-detection datasets without post-processing.
Takeaways & Limitations
Future evaluation is needed for semantic changes, other modalities such as SAR, image sequences, and training on larger datasets.
Abstract
from arXiv · showhide
This paper presents three fully convolutional neural network architectures which perform change detection using a pair of coregistered images. Most notably, we propose two Siamese extensions of fully convolutional networks which use heuristics about the current problem to achieve the best results in our tests on two open change detection datasets, using both RGB and multispectral images. We show that our system is able to learn from scratch using annotated change detection images. Our architectures achieve better performance than previously proposed methods, while being at least 500 times faster than related systems. This work is a step towards efficient processing of data from large scale Earth observation systems such as Copernicus or Landsat.
1. INTRODUCTION
Change detection assigns pixelwise change labels to coregistered images, supporting analyses of evolving land use and environmental conditions. The paper addresses the need for efficient learning from limited annotated datasets by proposing end-to-end fully convolutional architectures.
- Motivation: Change detection assigns a binary label to each pixel by comparing coregistered images acquired at different times.Positive labels indicate that the corresponding area changed between acquisitions.
- Motivation: Earth observation programs such as Copernicus and Landsat provide large image collections for supervised change-detection analysis.The available imagery motivates efficient methods for processing large-scale data.
- Motivation: A lack of large annotated change-detection datasets limits the complexity of models that can be trained.Pixelwise annotated datasets such as OSCD and Air Change remain available for supervised learning.
- Contribution: The paper presents three FCNN architectures trained end-to-end from scratch on available change-detection datasets for RGB and multispectral imagery.The architectures extend earlier end-to-end change-detection methods to a fully convolutional paradigm.
2. RELATED WORK
Change-detection methods have progressed from handcrafted pixel techniques to deep learning, but limited data has encouraged transfer learning and non-end-to-end approaches. This paper instead focuses on models that learn solely from change-detection data and support different image modalities.
- Research evolution: Change-detection research evolved from handcrafted pixel analysis through descriptors and simple machine learning toward deep learning.This progression followed broader developments in computer vision and image analysis.
- Transfer learning: Limited change-detection data led many methods to use networks pretrained on larger datasets for other problems.Transfer learning is presented as a common response to the scarcity of annotated data.
- Transfer learning: Pretraining can assume similarities between source and change-detection datasets and may not transfer directly to SAR or multispectral imagery.The cited RGB-pretrained networks are not directly applicable to the multispectral dataset discussed in the paper.
- Paper focus: The paper targets end-to-end models that learn solely from available change-detection data and can therefore be applied to available datasets.This focus avoids dependence on pretraining or transfer learning from other datasets.
3. PROPOSED APPROACH
The proposed approach converts patch-based change-detection ideas into fully convolutional encoder-decoder networks. It includes an early-fusion model and two Siamese variants that use shared-weight streams and different skip-connection combinations.
- Design goals: The fully convolutional architectures learn change detection end-to-end without pretraining or transfer learning and accept inputs of any size given sufficient memory.Moving from patches to a fully convolutional scheme improves inference speed and accuracy without significantly affecting training times.
- Architectural basis: The earlier Early Fusion model concatenated two image patches before processing them as different color channels, whereas Siamese branches initially processed images separately.Siamese branches used identical structure and shared parameters before merging.
- Skip connections: Skip connections link encoder and decoder layers at the same subsampling scale to combine abstract encoded information with earlier spatial details.This design supports accurate class predictions with precise output boundaries.
- FC-EF: FC-EF adapts U-Net as a shallower four-level encoder-decoder whose input concatenates the two images.It uses four max-pooling and four upsampling layers instead of the five levels in U-Net.
- Siamese variants: FC-Siam-conc and FC-Siam-diff split the encoders into shared-weight streams, differing by concatenated skip features versus absolute skip-feature differences.The difference heuristic is designed for detecting changes between the two images.
4. EXPERIMENTS
The experiments evaluate the proposed architectures on OSCD and Air Change datasets across multispectral and RGB settings, using quantitative and qualitative comparisons with prior methods. The fully convolutional models achieve strong detection results with inference below 0.1 seconds per image and substantial speed advantages.
- Datasets and setup: Two openly available datasets were evaluated: multispectral OSCD, RGB Air Change, and RGB-only OSCD inputs.Class weighting, geometric augmentation, and dropout were used during training.
- Datasets and setup: The evaluation reports change-class precision, recall, F1, and global precision when available, alongside comparisons with prior methods.OSCD comparisons include Early Fusion and Siamese methods; Air Change comparisons include DSCN, CXM, and SCCN.
- OSCD results: Under 0.1 s per image, the fully convolutional architectures predicted OSCD change maps far faster than the patch-based voting approach, which took several minutes.On OSCD, FC-EF and FC-Siam-diff achieved the best F1 scores, while FC-Siam-conc remained superior to patch-based approaches.
- Air Change results: On Air Change, all proposed architectures outperformed comparison methods in F1 for Szada/1, while FC-EF achieved the best F1 for Tiszadob/3.For Tiszadob/3, DSCN and SCCN outperformed the other proposed architectures.
- Interpretation: FC-Siam-diff appeared most suited to change detection, followed closely by FC-EF, reflecting explicit image comparison and difference-guided skip connections.The FC-EF architecture also achieved excellent results while learning these heuristics from training data.
- Practical significance: The reported speedup with no performance loss supports efficient processing of large Earth-observation data streams from programs such as Copernicus and Landsat.The authors describe this as a step toward accurate and fast worldwide monitoring.
5. CONCLUSION
The paper presents fully convolutional networks trained end-to-end from scratch that surpass prior change-detection methods in accuracy and inference speed. Future work includes testing semantic changes, other modalities, image sequences, and larger datasets.
- The three fully convolutional networks were trained end-to-end from scratch and surpassed the state of the art in accuracy and inference speed without post-processing.
- Future evaluation should test semantic-change detection, additional modalities such as SAR, and image sequences.
- Training on larger datasets is identified as another likely way to improve the networks.