Urban climate zone classification using convolutional neural network and ground-level images/
Xu, Guang
Urban climate zone classification using convolutional neural network and ground-level images/ - Sage, 2019. - Vol 43, issue 3, 2019 : (410-424 p.).
Urban climate risks have a wide range of impacts on the health of more than 50% of the world’s population, which is a critical issue relating to climate change. To support urban climate study and categorise different urban environments and their atmospheric impacts in a consistent way, the Local Climate Zone (LCZ) classification scheme has been developed. The World Urban Database and Access Portal Tools project aims to map the LCZ of cities across the globe. However, previous classification approaches based on satellite images have limitations regarding the characterisation of three-dimensional features such as building heights. This study aims to apply convolutional neural networks to classify LCZ types based on ground-level images, which can provide more detail of the urban environments. Validation results have shown an overall accuracy of 69.6%. The new method outperformed previous satellite-based studies for classifying the LCZ types Compact Mid-rise, Sparsely Built, Heavy Industry, and Bare Rock or Paved.
Urban climate,
Local Climate Zone,
convolutional neural network,
transfer learning,
Google Street View
Urban climate zone classification using convolutional neural network and ground-level images/ - Sage, 2019. - Vol 43, issue 3, 2019 : (410-424 p.).
Urban climate risks have a wide range of impacts on the health of more than 50% of the world’s population, which is a critical issue relating to climate change. To support urban climate study and categorise different urban environments and their atmospheric impacts in a consistent way, the Local Climate Zone (LCZ) classification scheme has been developed. The World Urban Database and Access Portal Tools project aims to map the LCZ of cities across the globe. However, previous classification approaches based on satellite images have limitations regarding the characterisation of three-dimensional features such as building heights. This study aims to apply convolutional neural networks to classify LCZ types based on ground-level images, which can provide more detail of the urban environments. Validation results have shown an overall accuracy of 69.6%. The new method outperformed previous satellite-based studies for classifying the LCZ types Compact Mid-rise, Sparsely Built, Heavy Industry, and Bare Rock or Paved.
Urban climate,
Local Climate Zone,
convolutional neural network,
transfer learning,
Google Street View