Developing High Quality Training Samples for Deep Learning Based Local Climate Classification in Korea

Minho Kim, Doyoung Jeong, Hyoungwoo Choi, Yongil KimarXiv 2020 · arXiv

Abstract

Local Climate Zone (LCZ) classification is a critical tool for monitoring urban heat islands and promoting sustainable urban development. However, global LCZ models often suffer from "domain adaptation" issues when applied to specific regions like South Korea, where architectural styles and vegetation cycles differ from global averages. This study presents a localized approach using a high-quality custom dataset and a Multi-Scale Convolutional Neural Network (MSCNN). By integrating local architectural data with satellite imagery, the proposed method achieves an overall accuracy of 83.88%, significantly outperforming state-of-the-art global models.

Introduction

As urban populations continue to grow, the Local Climate Zone (LCZ) framework has emerged as the gold standard for categorizing urban landscapes into 17 distinct classes based on physical structure and land cover. While global projects like WUDAPT and datasets like So2Sat provide a foundation for large-scale mapping, they often lack the precision required for regional applications. In Korea, for instance, the visual signature of a "compact mid-rise" building differs substantially from its European counterparts. This study addresses this gap by developing a localized training workflow tailored to the Korean urban context.

Methodology

The researchers moved away from "one-size-fits-all" global models, focusing instead on a data-rich, localized strategy involving three key components:

  1. Multi-Scale CNN (MSCNN): The team developed a specialized architecture featuring a "multi-scale layer." This layer processes input imagery through multiple kernel sizes simultaneously, allowing the network to capture both fine-grained architectural details and broader spatial contexts.
  2. Enhanced Labeling with MBI: To ensure ground-truth accuracy, the study integrated the Master Building Information (MBI) dataset from Korea’s Electronic Architectural Administration Information System (EAIS). This provided objective data on building heights and floor counts, which are notoriously difficult to estimate from 2D satellite imagery alone.
  3. Data Fusion: The model utilizes Sentinel-2 multispectral imagery (Bands 1-10) supplemented by the Normalized Difference Vegetation Index (NDVI) to better distinguish between natural land cover classes.
The architecture of the MSCNN featuring the multi-scale layer for capturing diverse spatial features.

Key Results

The custom MSCNN approach demonstrated a clear superiority over both traditional machine learning and global transfer learning methods. When tested on Korean urban data, the results were as follows:

  • MSCNN (Custom Data): 83.88% Overall Accuracy (Kappa: 64.99%)
  • Random Forest (Custom Data): 74.27% Overall Accuracy (Kappa: 41.00%)
  • So2Sat (Transfer Learning): 54.42% Overall Accuracy (Kappa: 48.60%)

The poor performance of the So2Sat model (54.42%) highlights the domain adaptation problem; global models struggle with local variations in crop fields and architectural signatures. The MSCNN was particularly adept at identifying "Compact High-rise" (LCZ 1) and "Open High-rise" (LCZ 4) areas, which are prevalent in Korean cities.

Quantitative performance comparison showing Overall Accuracy, Kappa coefficients, and F1-scores across all LCZ classes.

Discussion and Insights

The study reveals that local context trumps big data. Even a smaller, locally-optimized dataset can outperform massive global datasets if it accounts for regional architectural nuances. The use of MBI data acted as a "cheat sheet" during the labeling phase, effectively teaching the CNN to recognize visual signatures of building height that simpler models often confuse.

However, challenges remain. The models still struggle to distinguish between highly similar classes, such as "Compact Mid-rise" (LCZ 2) and "Compact Low-rise" (LCZ 3), where F1-scores remained below 50%. This suggests that visual spectral data alone may have reached its limit for these specific categories.

Conclusion

By combining multi-scale deep learning with high-quality local architectural data, this research provides a robust framework for regional LCZ mapping. Future work will focus on integrating 3D Digital Surface Models (DSMs) to provide explicit height information and scaling the sampling methodology to generate a comprehensive LCZ map for the entire Korean peninsula.

Summary updated

BibTeX
@misc{kim2020developing,
  author = {Minho Kim and Doyoung Jeong and Hyoungwoo Choi and Yongil Kim},
  title = {{Developing High Quality Training Samples for Deep Learning Based Local Climate Classification in Korea}},
  year = {2020},
  eprint = {2011.01436},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2011.01436},
}
Figure