Skip to content

lauraset/Super-resolution-building-height-estimation

Repository files navigation

A deep learning-based super-resolution method for building height estimation at 2.5 m spatial resolution in the Northern Hemisphere

Author: Yinxia Cao, Qihao Weng* | Paper link | Date: August 2024 | Journal: Remote Sensing of Environment

Dataset

1. pre-processed data

  • Download link in google drive, only with samples or Onedrive, full data. The total size is 2.72G
  • Unzip them to a path (e.g., data)
  • Split dataset into train/val/test set, see the directory data The specific spliting file is put in BH_dataset.py
  • Obtain the statistics of the dataset, see the directory datasteglobe The specifi file is stats_dataset_globe.py
  • Distribution of the dataset (45,000 samples) figure figure

2. the original data

  • original building height data: onedrive
  • original sentinel-1/2 data for sampling construction:
    They were downloaded from GEE for each image patch (640 x 640 m), and therefore there is no orginal data.Just see 1. the pre-processed data.

Method

figure

Setup environments

Necessary libraries were put in file requirements.txt

Training & Testing

python train.py

Pretrained weights and test results

See onedrive

  • weights of the super-resolution module: weights/realesrgan
  • weights of the proposed method for height estimation: weights/realesrgan_feature_aggre_weight_globe
  • Results on testing set: figure

Testing on 301 urban centers

  • Basic information
    For large metropolitan areas, the file name is xx_large, while for metropolitan areas, the file name is xx_metro. figure
  • Data and predicted results: see onedrive
    The mean and std of each urban center is put in datasetglobe/urbanarea_meanstd.xls and the results are put in: figure
  • Predicting by yourself
    • Download and put the origin data (S1&S2) in the current directory data/urban/input_data
      figure

    • Prepare xx_grid.shp as the prediction unit

      • Download WSF (world settlement footprint) 2019 at 10-m resolution as the valid built-up areas, see the official website
      • Clip WSF to the extent of each urban center
      • Convert the shapefile of urban centers into rasters
      • Split each urban center into grid with size of 640 x 640 m
      • Select the grid with wsf > 20 pixels
    • Predict all the urban centers

python predict_realesanet_feature_globe.py
  • Results:
    • Distribution figure
    • The mean and std of building height in each urban center figure

Predict other regions

  • Download sentinel-1 & sentinel-2 from the GEE
  • Download WSF

Other files

  • Download sentinel-1 & sentinel-2 from the GEE platform coming soon

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages