TL;DR
This paper presents a novel two-stage method for fine-grained building classification in rural areas using GF-7 satellite data. By combining supervised roof classification with unsupervised clustering of color and shape features, and incorporating a new template-based height correction for pitched roofs, the approach achieves high accuracy and provides detailed building type information crucial for disaster management.
What Problem Does This Paper Address?
Accurate building type information is vital for various applications, including disaster risk assessment, urban planning, and population studies. However, obtaining this information, especially in rural areas, is challenging due to the scarcity and outdated nature of existing data. Traditional methods like field surveys are time-consuming and labor-intensive. While remote sensing offers a solution, existing studies often focus on urban areas and struggle with the timeliness and spatial consistency of data, particularly for rural building footprints and heights. This paper aims to bridge this gap by developing a robust fine-grained building classification method specifically for rural environments, leveraging readily available, high-resolution satellite data.
Methods
The study employs a two-stage methodology:
- Supervised Roof Classification:
- Data Integration: Utilizes building footprints, building heights (derived from GF-7 stereo imagery), and multispectral information from GF-7 data.
- Classification Models: Compares Extreme Gradient Boosting (XGBoost) and Random Forest classifiers for roof type classification. XGBoost is chosen for its scalability, efficiency, and strong prediction performance.
- Feature Engineering: Incorporates spectral information (four multispectral bands), panchromatic reflectance, and geometric features derived from building footprints (area, length, width, aspect ratio, standard deviation of panchromatic reflectance).
- Height Correction for Pitched Roofs: A novel template-based height correction method is introduced. This method combines GF-7 derived building height with estimated roof slope angles (from street view photos) and building footprint geometry to improve height accuracy for pitched roofs, which are prone to stereo matching errors.
- Unsupervised Clustering for Fine-Grained Classification:
- K-Means Clustering: Applied to buildings within each identified roof type to further differentiate them based on color and shape features.
- Feature Spaces: Clustering is performed separately in color (multispectral bands, panchromatic standard deviation) and shape (area, length, width, aspect ratio) dimensions.
- Statistical Indicators: The Calinski-Harabasz index is used to determine the optimal number of clusters for each roof type and feature space.
- Code Generation: Fine-grained building types are represented by codes combining roof type, color cluster, shape cluster, and height classification (e.g., R1C2S3H1).
Datasets
- GF-7 Satellite Data: High-resolution multispectral imagery and stereo mapping data from the Gaofen-7 (GF-7) satellite, acquired on April 3, 2020. This data provides both spectral and height information.
- ICESat-2 LiDAR Data: Used for validating the extracted building heights.
- Street View Photos: Collected to estimate roof slope angles for the pitched roof height correction method.
- Study Area: Rural areas in Pingquan, Hebei Province, China, characterized by diverse topography and land cover.
Key Insights and Ideas
- GF-7 as a Unified Data Source: The study demonstrates the effectiveness of using a single satellite platform (GF-7) to acquire both high-resolution multispectral imagery and reliable building height information, simplifying data acquisition for rural areas.
- Template-Based Height Correction: The proposed template-based method significantly improves the accuracy of pitched roof height estimation, addressing a key limitation in using stereo imagery for such structures. This is crucial as pitched roofs are common in rural residential buildings.
- Two-Stage Classification Synergy: Combining supervised roof classification with unsupervised clustering allows for a hierarchical approach, first identifying broad roof categories and then refining them based on finer spectral and geometric details.
- Importance of Geometric Features: Beyond spectral information, geometric features derived from building footprints (like area, length, and width) are shown to be highly important for both roof classification and subsequent clustering.
- Rural Focus is Crucial: The paper highlights the urgent need for fine-grained building classification in rural areas due to their high vulnerability to natural disasters, a gap often overlooked in favor of urban studies.
Key Results
- Roof Type Classification: The XGBoost model achieved an overall roof type classification accuracy of 88.89%, outperforming the Random Forest model in overall performance.
- Height Correction: The template-based height correction method reduced the RMSE of pitched roof building heights from 2.28 m to 1.20 m.
- Fine-Grained Classification: A total of 41,160 buildings were classified into 140 categories. The most common building type, R1C1S3H1 (small single-storey buildings with light-colored pitched roofs), accounted for 18.17% of all classified buildings.
- Variable Importance: For XGBoost classification,
PanStd(standard deviation of panchromatic reflectance),area, andwidthwere identified as the most important variables. In cluster analysis, spectral band 3 and shape features likeareaandlengthwere most important.
Limitations and Future Work
- Irregularly Shaped Buildings: Buildings with footprints less than two-thirds of their minimum bounding rectangle were considered irregularly shaped and marked as unclassified, indicating a need for improvement in handling such cases.
- Vegetation and Dense Areas: The method's performance in areas with dense vegetation and more complex building arrangements (like southern China) needs to be tested. Higher resolution imagery might be beneficial in such scenarios.
- Deep Learning Exploration: The authors suggest exploring deep learning models, particularly those incorporating transformer architectures and attention mechanisms, for future work.
- Broader Application: The fine-grained building classification product is seen as a fundamental element for disaster management, providing a basis for vulnerability assessment, relocation planning, and emergency response.
Summary updated