Fully Convolutional Siamese Networks for Change Detection (1810.08462v1)

Abstract: 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.

• The paper presents three FCNN architectures that detect changes from multi-temporal imagery, outperforming conventional patch-based methods.
• The proposed models—FC-EF, FC-Siam-conc, and FC-Siam-diff—are trained end-to-end without pre-trained weights to effectively handle both RGB and multispectral data.
• Evaluations on OSCD and Air Change datasets show superior F1-scores and inference times under 0.1 seconds per image, enabling scalable real-time monitoring.

Fully Convolutional Siamese Networks for Change Detection

Introduction

This paper by Rodrigo Caye Daudt, Bertrand Le Saux, and Alexandre Boulch introduces three Fully Convolutional Neural Network (FCNN) architectures designed specifically for the task of Change Detection (CD) in Earth observation imagery. Leveraging both RGB and multispectral images, these architectures are engineered to process multi-temporal pairs of coregistered images to detect changes with improved accuracy and efficiency. CD is a critical area in remote sensing and Earth observation, addressing applications such as urban expansion monitoring, deforestation tracking, and environmental change analysis.

Background

Traditional change detection has witnessed notable shifts from manual pixel-level techniques to leveraging machine learning, and more recently deep learning frameworks. The paper acknowledges the lack of large annotated datasets as a limiting factor in CD, which makes end-to-end training approaches like those proposed here particularly valuable.

Prior methods typically rely on transfer learning from pre-trained networks, which may not generalize well to different sensor modalities. In contrast, the architectures presented here learn exclusively from CD-specific datasets, offering more adaptability and improved performance.

Proposed Architectures

The paper presents three FCNN architectures:

1. FC-EF (Fully Convolutional Early Fusion): This model is based on the U-Net architecture optimized for the CD task. The architecture employs skip connections to capture spatial details at various scales.
2. FC-Siam-conc (Fully Convolutional Siamese - Concatenation): This Siamese architecture processes each image in the pair through separate yet identical encoder networks with shared weights, followed by concatenation of their intermediate representations.
3. FC-Siam-diff (Fully Convolutional Siamese - Difference): This architecture takes the difference of the intermediate representations of the two images before concatenation, explicitly focusing on the differences between the images.

All models are trained end-to-end from scratch without pre-trained weights, thus bypassing the limitations inherent in transfer learning from large, non-related datasets.

Evaluation and Results

The architectures were evaluated on two publicly available datasets:

1. OSCD (Onera Satellite Change Detection):
   • The FCNN architectures significantly outperformed the previously proposed patch-based methods.
   • In terms of F1-score and global accuracy, FC-EF and FC-Siam-diff demonstrated superior performance.
   • The inference time for the FCNN methods was under 0.1 seconds per image, compared to several minutes for the patch-based systems.
2. Air Change Dataset:
   • The proposed methods were compared against state-of-the-art methods like DSCN, CXM, and SCCN.
   • On the Szada/1 and Tiszadob/3 cases, the FCNN methods, particularly FC-Siam-diff, yielded better F1-scores and inference speed, with the FC-EF architecture excelling in the Tiszadob/3 case.

Implications

The implications of this research are manifold:

• Operational Efficiency: The significant reduction in inference time demonstrates the potential for real-time analysis of Earth observation data streams.
• Scalability: These methods can be integrated into large-scale monitoring systems, such as those from Copernicus and Landsat, enabling rapid and accurate global change detection.
• Versatility: The ability to use these methods on both RGB and multispectral images without relying on pre-trained weights highlights their adaptability to various sensors and datasets.

Conclusion and Future Work

The paper sets a new benchmark in CD by presenting FCNN architectures that improve both accuracy and speed. Future directions could explore applying these architectures to semantic change detection, different imaging modalities like SAR, and temporal sequences analysis. Larger annotated datasets would further enhance the training and performance of these networks, broadening their applicability and reliability.

This research thus contributes a significant advancement in the efficiency and effectiveness of change detection methodologies, opening avenues for comprehensive, real-time environmental monitoring.