Abstract
High spatial resolution remote sensing imagery offers rich structural detail, but its spectral characteristics are less stable than traditional medium-low resolution data, often leading to poor classification results with pixel-based methods. This paper presents a multiscale spatial feature fusion approach. By employing wavelet transforms to compress neighborhood features and Support Vector Machines (SVM) for classification, the method dynamically selects the optimal scale for each pixel. Experimental results on QuickBird and IKONOS datasets demonstrate significant improvements in classification accuracy.
Introduction
Traditional remote sensing classification relies heavily on the spectral values of individual pixels. While effective for low-to-medium resolution data, this approach often fails for high-resolution imagery (e.g., QuickBird or IKONOS). In high-resolution data, "intra-class variance" is high—meaning pixels of the same land cover may look spectrally distinct due to shadows or textures—while "inter-class variance" decreases. This results in the "salt-and-pepper" effect, where classified maps appear noisy and fragmented. To address this, researchers must incorporate spatial context, though determining the appropriate neighborhood size (scale) for different objects remains a challenge.
Methodology
The proposed pipeline integrates multiscale spatial features through the following steps:
1. Dimensionality Reduction
To simplify the input, Principal Component Analysis (PCA) is performed on the multispectral bands. The first principal component, which contains the majority of the spatial and structural information, is extracted for further processing.
2. Multiscale Feature Extraction
For every pixel, the algorithm examines four nested neighborhood window sizes: $2\times2$, $4\times4$, $8\times8$, and $16\times16$. This allows the model to capture both fine-grained details and broader contextual patterns.
3. Wavelet-Based Compression
To prevent the feature vector from becoming computationally prohibitive, a Discrete Wavelet Transform (specifically the db3 wavelet) is used to compress information from the larger windows back into a $2\times2$ "root" feature set. This retains essential structural details while maintaining a uniform dimensionality across scales.
4. SVM Classification and Decision Fusion
An SVM with an RBF kernel is trained for each of the four scales. The final classification is determined by a scale selection factor, which evaluates neighborhood homogeneity and spectral variance to dynamically choose the most likely prediction for each specific pixel.
Experimental Results
The algorithm was evaluated on urban and suburban datasets from Beijing using QuickBird (2.44m resolution) and IKONOS (1m resolution) imagery.
- QuickBird Results: The proposed multiscale fusion achieved 88.6% accuracy (Kappa: 0.857), significantly outperforming traditional Maximum Likelihood Classification (75.6%) and single-pixel SVM (79.7%).
- IKONOS Results: The fusion method reached 82.5% accuracy, a substantial improvement over the 62.7% achieved by single-pixel methods on panchromatic data.
Key Insights
- Scale Trade-offs: Small scales ($2\times2$) excel at preserving sharp edges and narrow features like roads, while larger scales ($16\times16$) produce cleaner results in homogeneous areas.
- Dynamic Selection: The core innovation is the dynamic selection of scales, naturally favoring smaller scales for boundaries and larger scales for object interiors.
- Efficiency: Using wavelets as feature compressors provides the SVM with contextual summaries rather than raw pixel blocks, making the learning process more efficient.
Conclusion
By moving beyond simple pixel-based analysis to a multiscale spatial approach, this framework effectively eliminates salt-and-pepper noise while maintaining crisp boundaries for complex land features. Future work may explore extending this to multi-band spatial feature extraction to capture even greater spectral-spatial detail.
Summary updated