Vectorized rooftop area data for 90 cities in China

Zhixin Zhang, Zhen Qian, Teng Zhong, Min Chen, Kai Zhang, Yang Yue, Rui Zhu, Fan Zhang, Haoran Zhang, Fangzhuo Zhou, Jianing Yu, Bingyue Zhang, Guonian Lü, Jinyue YanScientific 2022 · DOI

TL;DR

This paper introduces a novel geospatial artificial intelligence (GeoAI) framework to generate vectorized rooftop area data for 90 cities across China using high-resolution remote sensing imagery. This dataset, validated with high accuracy, offers crucial insights for urban planning and sustainable development.

What Problem Does This Paper Address?

Accurate and up-to-date rooftop data is vital for urban planning, building energy modeling, and sustainable development. However, acquiring this data at a large scale is challenging. Traditional methods like 3D modeling via aerial photogrammetry are expensive, and conventional computer vision techniques struggle with the complexity and scale of urban environments. Existing open-access data sources often lack the necessary detail or are not readily available for many regions, particularly in developing countries. This paper aims to bridge this gap by providing a cost-effective and accurate method for generating vectorized rooftop data for a significant number of Chinese cities.

Methods

The core of this work is a geospatial artificial intelligence (GeoAI) framework designed for rooftop extraction. Key components include:

  • Data Preparation: This involves spatial stratified sampling informed by prior knowledge of urban land cover (from FROM-GLC30 data) to create balanced training and testing datasets. This strategy addresses the common issue of imbalanced classes (built-up vs. unbuilt areas). Preprocessing of Global Earth Satellite (GES) imagery includes gamma correction and contrast-limited adaptive histogram equalization to enhance image quality.
  • Deep Learning Model: The study utilizes the DeepLabV3+ model, a popular image semantic segmentation architecture. To improve performance, they integrated a composite loss function combining cross-entropy and dice loss to handle class imbalance and accelerate convergence.
  • Expansion Prediction: To mitigate issues with stitching artifacts at the boundaries of cropped image patches during prediction, an expansion prediction technique is employed. This involves expanding image patches before prediction and then extracting the central portion, ensuring seamless results.
  • Ensemble Learning: The final rooftop extraction model is an ensemble of two local optimal models (snapshots from specific training epochs), combined using a union ensemble rule. This approach balances prediction quality and speed.
  • Optimization: The model was fine-tuned using pre-trained weights from the MS COCO dataset and an Xception backbone. The AdamW optimizer and Cosine Annealing Warm Restarts algorithm were used for efficient training. The TensorRT framework was employed to accelerate inference speed.

Datasets

  • Global Earth Satellite (GES) imagery: High-resolution, open-access satellite imagery served as the primary data source for rooftop extraction. This imagery was downloaded via the Google Earth API.
  • FROM-GLC30: A 30-m spatial resolution global land cover dataset used to inform the spatial stratified sampling strategy.

The paper introduces a newly generated vectorized rooftop area dataset for 90 cities in China. This dataset is distinct from existing benchmarks due to its comprehensive coverage of numerous cities and its vectorized format.

Key Insights and Ideas

  • GeoAI for Urban Data: The successful application of a GeoAI framework demonstrates its potential for generating valuable urban geospatial data at scale, overcoming limitations of traditional methods.
  • Stratified Sampling is Crucial: The paper highlights the importance of spatial stratified sampling, informed by land cover data, for creating balanced datasets and improving model performance, especially when dealing with imbalanced classes like built-up vs. unbuilt areas.
  • Expansion Prediction for Seamless Results: The expansion prediction technique effectively addresses the common problem of stitching artifacts in large-scale image segmentation, leading to more visually coherent and accurate results.
  • Ensemble Models for Robustness: Combining multiple models through ensemble learning enhances the robustness and accuracy of the rooftop extraction, particularly when dealing with variations in image quality.
  • Data Quality Impacts Performance: The study acknowledges that variations in GES image quality across different regions can affect model performance, suggesting a need for careful consideration of data quality when applying the dataset.

Key Results

  • High Accuracy: The generated rooftop area dataset achieved an overall accuracy of 97.95% and an F1 score of 83.11% on a test set of 180 km².
  • City Tier Performance: The quality of rooftop extraction was generally better for higher-tier cities (Tier 1 and 2) compared to lower-tier cities (Tier 3 and 4), attributed to better GES image quality in more developed urban areas.
  • Dataset Size: The dataset comprises vectorized rooftop data for 90 cities in China, available in ESRI Shapefile format. The total size of the original and simplified versions is 118 GB and 21.3 GB, respectively.
  • Spatial Conformity: The generated rooftop areas conform to urban morphological characteristics and reflect urbanization levels, making them suitable for urban planning and analysis.

Limitations and Future Work

  • Image Quality Variations: The authors acknowledge that variations in GES image quality (due to sensor, time, and atmospheric conditions) can impact model performance and generalization. This leads to performance differences across city tiers.
  • Position Offset: For a few high-rise buildings, a slight position offset (approximately 20 m) was observed between extracted rooftops and ground truth, attributed to potential issues with image orthorectification. While negligible for large-scale estimations, it might be a consideration for specific high-resolution architectural applications.
  • Generalization to Other Regions: While the method shows strong robustness, extending it to regions outside China might require fine-tuning the model with new feature samples to enhance applicability.
  • Future Work: The paper suggests that the dataset can be updated regularly to reflect ongoing urbanization. Further research could explore the impact of regional quality differences on specific applications and investigate methods to further improve accuracy in challenging areas.

Summary updated

Abstract

Reliable information on building rooftops is crucial for utilizing limited urban space effectively. In recent decades, the demand for accurate and up-to-date data on the areas of rooftops on a large-scale is increasing. However, obtaining these data is challenging due to the limited capability of conventional computer vision methods and the high cost of 3D modeling involving aerial photogrammetry. In this study, a geospatial artificial intelligence framework is presented to obtain data for rooftops using high-resolution open-access remote sensing imagery. This framework is used to generate vectorized data for rooftops in 90 cities in China. The data was validated on test samples of 180 km<sup>2</sup> across different regions with spatial resolution, overall accuracy, and F1 score of 1 m, 97.95%, and 83.11%, respectively. In addition, the generated rooftop area conforms to the urban morphological characteristics and reflects urbanization level. These results demonstrate that the generated dataset can be used for data support and decision-making that can facilitate sustainable urban development effectively.
BibTeX
@inproceedings{zhang2022vectorized,
  author = {Zhixin Zhang and Zhen Qian and Teng Zhong and Min Chen and Kai Zhang and Yang Yue and Rui Zhu and Fan Zhang and Haoran Zhang and Fangzhuo Zhou and Jianing Yu and Bingyue Zhang and Guonian Lü and Jinyue Yan},
  title = {{Vectorized rooftop area data for 90 cities in China}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41597-022-01168-x},
}
Figure