Local Climate Zone Mapping as Remote Sensing Scene Classification Using Deep Learning: A Case Study of Metropolitan China

Shengjie Liu, Qian Shi

Email: skrisliu@gmail.com

This paper is published in ISPRS Journal of Photogrammetry and Remote Sensing. [Paper]

Preprint via this link.

This is a GitHub repo at github.com/skrisliu/lcz.

---

New Data Released (2020-08-18)!

We've generated a new product of the whole Pearl River Delta region (9+2 cities), available here

New Features

• Used the Sentinel-2 composite images from 2018-2019 to generate the LCZ map, so the mosaic effect is reduced to a minimum.
• Included the whole administrative area of the following cities: Guangzhou, Foshan, Zhaoqing, Shenzhen, Dongguan, Huizhou, Zhongshan, Jiangmen, Zhuhai, Macau, and Hong Kong (广州，佛山，肇庆，深圳，东莞，惠州，中山，江门，珠海，澳门，香港)

Pearl River Delta (The Greater Bay Area, old version)

Tiff file is available here (only core urban areas included)

---

Training and Testing Data

Download Training and Testing Data from Google Drive or email me

Region	City
The Greater Bay Area	Guangzhou, Foshan, Shenzhen, Dongguan, Huizhou, Zhuhai, Zhongshan, Jiangmen, Macau, Hong Kong 广州，佛山，深圳，东莞，惠州，珠海，中山，江门，澳门，香港
The Shanghai Metropolis	Shanghai, Hangzhou, Shaoxing 上海，杭州，绍兴
The Beijing Metropolis	Beijing, Tianjin, Tangshan 北京，天津，及部分唐山

Load a pretrained model

Download Pretrained Model from Google Drive.

The size of input is 64*64 with 10 channels

The channel order is (channel & Sentinel-2 band):[0,1,2,3,4,5,6,7,8,9] = [2,3,4,5,6,7,8,11,12,8A]

Please use the following normalization when applying the model.
The subimage should be 64*64.

import keras
modelfile = 'modelpath'
model = keras.models.load_model(modelfile)
data = np.float32(data)
data = data/5000.0
for subimage in data:
	model.predict(subimage)