Unsupervised Deep Slow Feature Analysis for Change Detection in Multi-Temporal Remote Sensing Images (1812.00645v2)

Abstract: Change detection has been a hotspot in remote sensing technology for a long time. With the increasing availability of multi-temporal remote sensing images, numerous change detection algorithms have been proposed. Among these methods, image transformation methods with feature extraction and mapping could effectively highlight the changed information and thus has better change detection performance. However, changes of multi-temporal images are usually complex, existing methods are not effective enough. In recent years, deep network has shown its brilliant performance in many fields including feature extraction and projection. Therefore, in this paper, based on deep network and slow feature analysis (SFA) theory, we proposed a new change detection algorithm for multi-temporal remotes sensing images called Deep Slow Feature Analysis (DSFA). In DSFA model, two symmetric deep networks are utilized for projecting the input data of bi-temporal imagery. Then, the SFA module is deployed to suppress the unchanged components and highlight the changed components of the transformed features. The CVA pre-detection is employed to find unchanged pixels with high confidence as training samples. Finally, the change intensity is calculated with chi-square distance and the changes are determined by threshold algorithms. The experiments are performed on two real-world datasets and a public hyperspectral dataset. The visual comparison and quantitative evaluation have both shown that DSFA could outperform the other state-of-the-art algorithms, including other SFA-based and deep learning methods.

• The paper introduces DSFA, an unsupervised approach combining deep networks with slow feature analysis for robust change detection.
• It leverages symmetric deep networks and change vector analysis to extract distinct features from bi-temporal remote sensing data.
• The method achieves higher overall accuracy, Kappa coefficient, and F1 scores, demonstrating its effectiveness over traditional techniques.

Overview of "Unsupervised Deep Slow Feature Analysis for Change Detection in Multi-Temporal Remote Sensing Images"

The paper "Unsupervised Deep Slow Feature Analysis for Change Detection in Multi-Temporal Remote Sensing Images" proposes a novel approach for change detection by leveraging deep learning frameworks and slow feature analysis (SFA). The core innovation of this work is the development of an unsupervised algorithm called Deep Slow Feature Analysis (DSFA), which aims to improve the performance of change detection in remote sensing by addressing the complexities and variations inherent in multi-temporal datasets.

Main Contributions

The paper introduces DSFA, combining deep networks with the SFA theory, to enhance feature extraction and change detection capabilities. Specifically, the model utilizes symmetric deep networks to transform bi-temporal image data into a robust feature space. These transformations allow the model to distinguish between changed and unchanged features effectively.

Key contributions include:

• Integration of Deep Networks with SFA: The DSFA algorithm applies deep networks to form a non-linear projection of the input data. Unlike traditional linear models, these networks can better capture complex patterns and variations, making them more suited for handling diverse remote sensing datasets.
• SFA Module for Change Highlighting: The role of the SFA module is critical in DSFA. By reducing the unchanged components in the projected features, it enhances change detection. Importantly, the approach is data-driven and does not require pre-labeled data, positioning it as an unsupervised method that adapts well to varying conditions without needing extensive labeled samples.
• Pre-Detection via Change Vector Analysis (CVA): CVA is employed to pre-identify unchanged pixels which serve as high-confidence training samples, facilitating robust model training.
• Chi-square Distance for Change Intensity Calculation: A chi-square distance measure is utilized to calculate change intensity, ensuring sensitivity to variations across feature bands, aiding in precise threshold determination for change detection.

Numerical Results

The DSFA shows notable performance improvements across several metrics when tested on datasets including Taizhou and Nanjing ETM, and a public hyperspectral dataset. Experiments demonstrate that DSFA consistently outperforms existing methods:

• Overall Accuracy (OA): DSFA achieves higher OA, indicating precise delineation of changed and unchanged regions.
• Kappa Coefficient and F1 Score: These metrics further confirm the method's ability to reduce false positives and false negatives relative to other state-of-the-art algorithms.

Implications and Future Directions

DSFA marks a significant step in the enhancement of remote sensing change detection by addressing the limitations of both conventional SFA models and purely linear approaches. The adaptation of deep networks enables greater flexibility in processing various types and modalities of remote sensing data. The unsupervised nature of the approach makes it suitable for vast area applications where labeled data is scarce.

Theoretically, this work extends the applicability of SFA into higher-dimensional and more complex feature spaces achieved via deep learning, which could open avenues for further research in domain adaptation and transfer learning. Practically, its implementation could be crucial in environmental monitoring, disaster response, urban planning, and other areas reliant on satellite imagery.

For future developments, the integration of additional contextual or ancillary data and further exploration into model scalability and real-time application could be valuable. Additionally, extending DSFA to handle multi-class change detection might enhance its utility in more complex operational settings, expanding beyond simple binary classification tasks.