Abstract
Change detection (CD) is an important application of remote sensing, which provides information about land cover changes on the Earth's surface. Hyperspectral image (HSI) can show more spectral information, which greatly improves the ability of remote sensing to identify change features. The challenge is how to overcome the scarcity of labeled samples and extract the change information of high-dimensional spectra in HSI. To solve the previous problem, a semi-supervised CD with multilayer cascade screening strategy (MCS$^4$CD) that uses both the spatial information and active learning is proposed to select highly reliable unlabeled samples to increase the training sets. The MCS$^4$CD method can effectively use unlabeled samples to improve accuracy. Additionally, a subspace CD method based on iterative slow feature analysis is designed to extract the most temporally invariant component from the high-dimensional space. Experimental results on four hyperspectral datasets show that with a small number of labeled samples, the proposed method achieves a much better performance than existing CD methods.
Introduction
Change detection (CD) is a critical task in remote sensing, involving the comparison of satellite images of the same location taken at different times to identify environmental or urban shifts. Hyperspectral images (HSI) are the gold standard for this task because they capture hundreds of narrow spectral bands, providing a detailed "spectral fingerprint" of the surface. However, HSI processing faces two major hurdles: the "curse of dimensionality" due to high-dimensional data redundancy, and the scarcity of labeled samples, as human verification is often impractical in remote or disaster-stricken areas.
Methods
To address these challenges, the researchers propose a two-stage framework combining subspace feature extraction with a novel screening strategy.
Subspace Iterative Slow Feature Analysis (SISFA)
First, the authors introduce SISFA to handle high-dimensional data. Standard Slow Feature Analysis (SFA) identifies transformations where unchanged pixels remain constant over time. SISFA improves this by projecting data into a lower-dimensional subspace first, removing noise and spectral redundancy to better isolate invariant features.
Multilayer Cascade Screening Strategy (MCS$^4$CD)
The core of the semi-supervised approach is the MCS$^4$CD strategy, which acts as a high-precision filter to intelligently select reliable unlabeled samples to augment the training set.
The screening process consists of three layers:
- Initial Prediction: A Support Vector Machine (SVM) provides initial labels for unlabeled data.
- Spatial Screening: A "Circular Neighborhood" check ensures a pixel's label matches its neighbors, leveraging the First Law of Geography to filter out noise.
- Active Learning: A "Breaking Ties" strategy identifies the most informative, uncertain samples to refine the model's decision boundaries.
Results and Discussion
The model was evaluated on five hyperspectral datasets, including agricultural (Hermiston, USA) and urban/river scenes (River, China, Xiong County).
Key findings include:
- Superior Accuracy: MCS$^4$CD consistently outperformed traditional methods like MAD and ISFA. For instance, on the Hermiston dataset, it achieved an Overall Accuracy (OA) of 97.84%, significantly higher than the 88.17% achieved by standard ISFA.
- Label Efficiency: The semi-supervised approach proved highly effective with small sample sizes, achieving high performance with as few as 20 labeled samples per class.
- Robustness: In complex urban environments like Xiong County, the method maintained an OA of 89.19%, effectively suppressing the "salt-and-pepper" noise that typically plagues unsupervised methods.
Conclusion
The proposed MCS$^4$CD framework demonstrates that combining subspace feature extraction with a spatial-aware active learning screening process can overcome the limitations of labeled data scarcity in hyperspectral change detection. Future work will focus on incorporating nonlinear transformations and more advanced spatial constraints to further improve performance in complex, fragmented urban landscapes.
Summary updated