Home

Awesome

Codes and data used in the paper "Large-scale Automatic Identification of Urban Vacant Land Using Semantic Segmentation of High-Resolution Remote Sensing Images". https://doi.org/10.1016/j.landurbplan.2022.104384

The codes are based on a deep learning approach, using the semantic segmentation network DeepLabv3.

Code and data files: RS_dataset/ # RAW DATA AND OUTPUTS train_raw/ # store all cities' raw images and labels for training -BJ_01_map.jpg # eg for filename of image: {city}_{ID}map.jpg -BJ_01_label.jpg # eg for filename of label: {city}{ID}_label.jpg train_input/ # training images and labels of the current city model, # produced by -train.py test_raw/ # store all cities' raw images for test and prediction -BJ_test01_map.jpg
test_output/ # output predicton results, prodeced by -predict.py

data_process/ # DATA PROCESSING new_dataset/ # image and label patches, produced by codes -cut_img_train.py # code for splitting train images and labels -cut_img_test.py # code for splitting test images -merge_img_test.py # code for merging test model outputs -post_process.py # code for optimizing model outputs -train_val_txt.py # code for dividing training and validation sets -utils.py # code for generating datasets and defining evaluation metrics

model/ # DEEP LEARNING MODEL best_model/ # store the best deeplabv3 models for prediction -DeeplabV3_xxx.pth # 4 models: SZ, BJ, LZ, and Hybrid, used in the paper deeplabv3/ # code and pretrained backbones of deeplabv3 model/ # code for deeplabv3 model -deeplabv3.py # main code -resnet.py # encoder -aspp.py # decoder pretrained_models/resnet/ # pretrained resnet backbones -resnetxxx.pth -DeeplabV3_xxx.pth # the best model in the training process

log/ # MODEL TRAINING LOG csv/ # csv files -Log.log # text file

-EDA.py # main code for exploring the UVL rate of the training dataset, # which relates to the weighted loss function. -train.py # main code for model training -predict.py # main code for prediction

How to run the code: -Environment:

  1. Python 3.7
  2. PyTorch torch==1.4.0+cu92

-To train:

  1. Put raw images and labels in the folder "RS_dataset/train_raw/" following the filename rules above.
  2. Run -train.py for model training. Select city. Adjust args and settings as you like.
  3. Find -DeeplabV3_xxx.pth under the folder "model/", which is the best trained model.

-To predict:

  1. Put raw images in the folder "RS_dataset/test_raw/" following the filename rules above.
  2. Run -predict.py. Select model and city. Adjust args and settings as you like.
  3. Find results under the folder "RS_dataset/test_output/".

Training tips:

  1. Use -EDA.py to explore the UVL rate of the training dataset and use [rate, 1-rate] as the weights of the weighted loss function.
  2. The spatial resolution of the RS image is recommended to be around 1.6 m. See Appendix B of the paper for explanation.
  3. The default deeplabv3 model in the code uses resnet18 as backbone, which is recommended. See Appendix C for explanation.

The models and RS_dataset can be found in https://www.beijingcitylab.com/projects-1/56-urban-vacancies/ https://cloud.tsinghua.edu.cn/f/6a1437d2478f4727940a/