TL;DR
This paper pioneers global-scale Local Climate Zone (LCZ) classification using Sentinel-1 dual-polarization radar data. While achieving satisfactory accuracy remains a challenge, the study highlights the potential of SAR data and identifies key features, particularly morphological and VH-polarized ones, for this task.
What Problem Does This Paper Address?
The Local Climate Zone (LCZ) system categorizes urban areas into 17 classes based on their physical structure and land cover, offering valuable insights into urban heat islands, economic status, and population distribution. While previous LCZ classification efforts have focused on individual cities using optical remote sensing data, a global-scale classification is highly desirable for broader studies. However, achieving this global scale presents significant challenges, including the need for robust classification models that can generalize across diverse urban environments and the difficulty of obtaining sufficient, representative training data. This paper aims to address these challenges by exploring the use of Sentinel-1 dual-polarization Synthetic Aperture Radar (SAR) data for global LCZ classification and investigating various feature extraction techniques to improve classification performance and generalization.
Methods
The study employs a multi-faceted approach to feature extraction and classification:
- Feature Extraction: Four main categories of features were extracted from Sentinel-1 dual-pol SAR data:
- Polarimetric Features (Pol-Baseline): Intensities of VV and VH channels, normalized coherence, and relative phase between VV and VH.
- Local Statistical Features (Stat-Feature): Statistical parameters (max, min, mean, std, median) derived from local patches (10x10 pixels) of the polarimetric features.
- Texture Features (GLCM-Feature): Texture information extracted using the Gray-Level Co-occurrence Matrix (GLCM) from VV and VH intensity images.
- Mathematical Morphological Features (MP-Feature): Features derived from morphological profiles, utilizing opening and closing operations with varying structuring element sizes.
- Classification: A Canonical Correlation Forest (CCF) classifier was chosen. CCF is an advanced random forest method known for its generalization capabilities and ability to analyze feature importance. It projects data into a latent space where features are maximally correlated, aiming to find more natural class boundaries.
Datasets
- Sentinel-1 Dual-Pol SAR Data: Level-1 Interferometric Wide swath (IW) SLC products were used, providing VV and VH polarization channels.
- Ground Truth Data: Manually delineated LCZ polygons were created for 29 cities across various continents (excluding Antarctica). These cities were selected based on high population density and geographical diversity to ensure transcultural, transnational, and cross-environmental representation.
- Digital Surface Model (DSM): The SRTM dataset was used for terrain correction during data preprocessing.
Key Insights and Ideas
- SAR Data for LCZ: Sentinel-1 dual-pol SAR data, particularly its polarimetric and spatial-contextual information, shows promise for global LCZ classification, even though current accuracies are not yet satisfactory.
- Morphological Features Shine: Mathematical morphological features (MP-Feature) proved to be highly effective, significantly improving classification accuracy when combined with local statistical features. This highlights the importance of capturing the spatial structure of urban areas for LCZ classification.
- VH Polarization Dominance: Features derived from the VH polarization channel were found to be the most important contributors to the classification results, suggesting its sensitivity to urban structures.
- Coherence Matters: The coherence between VV and VH channels, often overlooked, also contributed positively to the classification, while relative phase information provided minimal benefit and even degraded performance.
- GLCM Limitations: GLCM-based texture features, while informative in other SAR applications, were found to be unsuitable for global LCZ classification due to their dependency on local statistical distributions, limiting their generalization capability across different cities and acquisition conditions.
Key Results
- The combination of local statistical features and morphological features (specifically, MP-Feature-F, i.e., with speckle filtering) yielded the best overall accuracy (OA) of 61.8%. This was a significant improvement of 3% over the second-best feature combination.
- The Stat-Feature alone outperformed the Pol-Feature by 5.68% in OA, demonstrating the value of incorporating spatial neighborhood information.
- Morphological profiles improved classification accuracy by 3% in OA compared to using only local statistical features.
- The study found that features related to VH polarization were among the top eight most important features, indicating their strong contribution.
Limitations and Future Work
- Accuracy Limitations: The paper acknowledges that the classification accuracy is "not yet satisfactory" for a global scale, with many urban classes being difficult to distinguish and often confused with "large low-rise."
- GLCM Unsuitability: The GLCM texture feature was found to be unsuitable for global LCZ classification due to its lack of generalization capability.
- Future Directions:
- Investigate deep learning strategies for automatic feature extraction and selection.
- Explore the optimal neighborhood size for remote sensing data in LCZ classification.
- Incorporate both radar and optical remote sensing data for data fusion.
- Develop solutions with stronger generalization capabilities for global LCZ classification.
Summary updated