Improving the local climate zone classification with building height, imperviousness, and machine learning for urban...

Area of Science:

• Environmental Science
• Remote Sensing
• Urban Climatology

Background:

• Local Climate Zone (LCZ) classification is crucial for urban heat island and climate studies.
• Current LCZ methods lack integration of vital urban auxiliary GIS data like building height and imperviousness, limiting accuracy and utility.
• Existing frameworks, such as WUDAPT, can be enhanced with additional datasets.

Purpose of the Study:

• To systematically compare and evaluate machine and deep learning methods for LCZ classification using a hybrid GIS and remote sensing approach.
• To assess the impact of incorporating building height and imperviousness data on LCZ classification accuracy.
• To identify optimal classification strategies and dominant features for improved urban climate studies.

Main Methods:

• A hybrid GIS- and remote sensing imagery-based framework was employed.
• Machine learning (Random Forest - RF) and deep learning (Convolution Neural Network - CNN) classifiers were systematically compared.
• Auxiliary datasets (building height, imperviousness) and spectral bands (near-infrared, thermal infrared) were analyzed for their contribution to classification accuracy.

Main Results:

• The Convolution Neural Network (CNN) achieved high accuracy but at the cost of spatial resolution due to multi-pixel input.
• The Random Forest (RF) classifier demonstrated superior performance among single-pixel classifiers.
• Incorporating building height data improved high- and mid-rise LCZ class accuracy in RF, while imperviousness data enhanced low-rise class accuracy.
• Auxiliary datasets were found to dominate RF classification accuracy, whereas spectral bands were key for CNN.

Conclusions:

• The integration of building height and imperviousness data significantly enhances the accuracy of LCZ classification.
• The Random Forest classifier, augmented with auxiliary GIS data, offers a robust approach for single-pixel LCZ mapping.
• The proposed framework provides an improvable and adaptable method for generating accurate LCZ maps for urban modeling across different cities.