[Paper Review] Developing High Quality Training Samples for Deep Learning Based Local Climate Zone Classification in Korea

Minho Kim, Doyoung Jeong+2 more|arXiv (Cornell University)|Nov 3, 2020

Urban Heat Island Mitigation4 citations

TL;DR

This study develops high-quality, custom Local Climate Zone (LCZ) training samples for major Korean cities using Sentinel-2 imagery, NDVI, and building data from the EAIS, then applies a multi-scale convolutional neural network (MSCNN) to achieve 83.88% overall classification accuracy—significantly outperforming both random forest and transfer learning from the global So2Sat dataset.

ABSTRACT

Two out of three people will be living in urban areas by 2050, as projected by the United Nations, emphasizing the need for sustainable urban development and monitoring. Common urban footprint data provide high-resolution city extents but lack essential information on the distribution, pattern, and characteristics. The Local Climate Zone (LCZ) offers an efficient and standardized framework that can delineate the internal structure and characteristics of urban areas. Global-scale LCZ mapping has been explored, but are limited by low accuracy, variable labeling quality, or domain adaptation challenges. Instead, this study developed a custom LCZ data to map key Korean cities using a multi-scale convolutional neural network. Results demonstrated that using a novel, custom LCZ data with deep learning can generate more accurate LCZ map results compared to conventional community-based LCZ mapping with machine learning as well as transfer learning of the global So2Sat dataset.

Motivation & Objective

• To address the limitations of low accuracy and poor generalization in global LCZ mapping, particularly for regional urban contexts like Korea.
• To develop a high-quality, region-specific LCZ training dataset for major Korean cities using multi-source remote sensing and GIS data.
• To evaluate the effectiveness of custom training data versus transfer learning from global datasets like So2Sat in deep learning-based LCZ classification.
• To improve the accuracy of LCZ mapping in Korea by leveraging multi-scale convolutional neural networks (MSCNN) on curated training samples.
• To establish a foundation for national-scale LCZ mapping in Korea using data-driven, high-resolution methods.

Proposed method

• Constructed 32×32 pixel image patches from Sentinel-2 mosaics, using NDVI for natural cover and MBI data for built-up cover to guide LCZ labeling.
• Generated LCZ labels by placing centroids at representative points and validated them using Google Earth and Street View to ensure accuracy and completeness.
• Employed a multi-scale convolutional neural network (MSCNN) with three parallel convolutional layers (kernels of 3, 5, and 7 pixels) to capture multi-scale spatial features.
• Applied data augmentation and class reweighting to mitigate class imbalance in the training samples.
• Trained the MSCNN using Adam optimizer, batch size 96, initial learning rate 0.002, and early stopping after 15 epochs.
• Conducted transfer learning using the So2Sat dataset by freezing the ResNet backbone and adding two new fully-connected layers for fine-tuning.

Experimental results

Research questions

• RQ1 Can custom, high-quality LCZ training samples for Korea significantly improve classification accuracy compared to community-based or global transfer learning approaches?
• RQ2 How does a multi-scale CNN (MSCNN) perform in classifying heterogeneous urban LCZs compared to traditional random forest models?
• RQ3 To what extent does transfer learning from the So2Sat global dataset generalize to Korean urban environments, given regional differences in land cover and seasonal patterns?
• RQ4 How do class-specific F1-scores vary across LCZ types, and which classes remain challenging despite advanced deep learning models?
• RQ5 Can integration of building height and GIS data enhance the precision of LCZ labeling and classification in complex urban areas?

Key findings

• The MSCNN model trained on custom Korean LCZ data achieved an overall accuracy of 83.88%, significantly outperforming the random forest model (74.27%) and So2Sat transfer learning (54.42%).
• The Kappa statistic for the MSCNN model reached 64.99%, indicating substantial agreement beyond chance, compared to 41.00% for RF and 48.60% for So2Sat.
• MSCNN demonstrated superior performance in classifying built-up LCZs (e.g., LCZ 1 and LCZ 4), especially in dense urban and suburban zones.
• The So2Sat transfer learning approach struggled with natural cover classes (e.g., LCZ 2, 3, 5, 6), showing confusion likely due to seasonal and regional differences in vegetation patterns.
• Class-wise F1-scores below 50% were observed for LCZ 2 vs. 3 and LCZ 5 vs. 6, indicating persistent challenges in distinguishing similar urban and vegetated zones.
• The study confirms that region-specific, high-quality training data is more effective than global transfer learning for local-scale LCZ mapping in Korea.