Local Climate Zone Classification Using a Multi-Scale, Multi-Level Attention Network

Authors: Minho Kim, Doyoung Jeong, Yongil Kim

The code is provided for the Multi-Scale, Multi-Level Attention Network (MSMLA) proposed in Local climate zone classification using a multi-scale, multi-level attention network accepted in ISPRS J.

The model is shown below:

All experiments were trained from scratch and were performed using an Intel Core i7-6700 CPU at 3.40 GHz and an NVIDIA GeForce RTX 2070 Super Graphics Processor Unit (GPU) with 8 GB of memory. Python 3.7.9 was used with Tensorflow 2.3.0. For training hyperparameters, an early stop of 15 epochs, a learning rate of 0.002, and a decay factor of 0.004 were used. The adaptive moment estimation (adam) optimizer was chosen to minimize the cross-entropy loss function. Filter weights were initialized using “He normal” initialization.

Requirements

python=3.7.9
tensorflow-gpu=2.3
tensorflow-addons
scikit-learn
matplotlib
pandas
albumentations (pip)

Usage

Install a new conda environment

$ conda env create --name msmla --file environment.yml

Activate the new environment and navigate to the "src" folder

$ conda activate msmla 
$ cd src

Download train, test, validation sample .h5 files. The data_link.rtf text file in the LCZ_MSMLA/data folder contains a download link. Make sure to move the .h5 files into the LCZ_MSMLA/data folder.
Run the train.py code

$ python train.py --path /YOUR/LOCAL/DIRECTORY/LCZ_MSMLA # Only change /YOUR/LOCAL/DIRECTORY without " "

LCZ Classification

Classification accuracy results of benchmark and proposed models using “All + Ancillary” (Sentinel-2 10 bands, OSM, DSM, NLCM) data with 48x48 pixel patches. Please refer to the paper for more results and details.

Model	Trainable Parameters	OA (%)	WA (%)	OA_BU (%)	OA_N (%)	F1 (%)
Sen2LCZ	793,348	81.0	73.9	54.4	84.8	78.7
LCZNet	3,181,809	83.0	79.8	63.8	86.5	82.2
CNN	235,105	83.4	78.1	61.1	89.0	82.2
MSMLA-18^*	181,867	84.4	80.4	64.7	89.4	83.5
MSMLA-50^*	808,913	87.1	85.0	72.4	91.8	86.5

^* Proposed models

🔥 State-of-the-art, Deep Learning-based LCZ Classification Models

[Sen2LCZ-Net] Qiu, C., Tong, X., Schmitt, M., Bechtel, B., & Zhu, X. X. (2020). Multilevel feature fusion-based CNN for local climate zone classification from sentinel-2 images: Benchmark results on the So2Sat LCZ42 dataset. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 13, 2793-2806.
[LCZNet] Liu, S., & Shi, Q. (2020). Local climate zone mapping as remote sensing scene classification using deep learning: A case study of metropolitan China. ISPRS Journal of Photogrammetry and Remote Sensing, 164, 229-242.
[CNN] Rosentreter, J., Hagensieker, R., & Waske, B. (2020). Towards large-scale mapping of local climate zones using multitemporal Sentinel 2 data and convolutional neural networks. Remote Sensing of Environment, 237, 111472.

Citation

Please cite the journal paper if this code is useful and helpful for your research.

@article{kim2021local,
  title={Local climate zone classification using a multi-scale, multi-level attention network},
  author={Kim, Minho and Jeong, Doyoung and Kim, Yongil},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={181C},
  pages={345--366},
  year={2021},
  publisher={Elsevier}
}

minhokim93 / LCZ_MSMLA