Mapping Sandy Beaches using Deep Learning

Mapping Australian sandy beaches using image segmentation deep learning method based on U-Net architecture.

Installation

git clone https://bitbucket.csiro.au/scm/~yon012/sandybeachmapping_dl.git
pip install -r requirements.txt

Quick Start

Download data: See Spatial processing and downloading of data

Run deep learning pipeline:

0. Setup configs: config/config.yaml
1. Run training: $ python -m sandybeachmapping_dl.main
2. Run evaluation: $ python -m sandybeachmapping_dl.main --config-name='config_eval.yaml'

Usage

Data folder layout and filename convention

• Folder: <data_dir>
• Filename convention: <IMAGE_SOURCE>_<AUS_STATE>_<IMAGE_TYPE>_FID.tif
  • <IMAGE_SOURCE>: Image source, e.g., from Bing
  • <AUS_STATE>: Australian state
  • <IMAGE_TYPE>: Either image, mask, or maskpred

Spatial processing and downloading of data

{R} scripts to tile and download images for input into {python} ML code, and then convert ML raster output to polygon vectors are found in the folder coast_tiling_extraction. The extraction scripts can be run for each Australian State (e.g. VIC).

0. Not used download_and_Tile_beach_images.R tile the region of a OSM beach into n x m tiles covering the entire OSM beach id polygon.
1. tile_beach_images.R define 512 x 512 image tiles along the coastline of a Australian state (e.g. VIC). The tiles overlap, but tiles with more than 5 overlaps are removed (rs > thr).
2. download_tile_images.R download the aerial images for each 512 x 512 tile in previous step.
3. mask_tile_images.R rasterise the binary "mask" of the OSM beach polygons for each image tile in the previous step.
   • Not used mask_tile_images_BeachCliffsSea.R make a classified mask to include a cliff and beach mask.
4. polygonise_ml_output_maskpred.R convert the ML output rasters to polygon vectors, to compare to OSM beaches.

Output files from main script

• Select test set to evaluate in configs: test_imageset
  • mask: Test images with masks
  • mask+nomask: Test images with masks and images without masks
  • all: All images, including from training, validation and testing. Outputs will be in <maskpred_subdir>_all.
• <output_dir>/<maskpred_subdir>/: GeoTiff of predicted masks
• <output_dir>/<maskpred_subdir>_polygons.shp.zip: Shapefile of predicted masks
• In <modelout_dir>/:
  • model.pth: Trained model
  • loss_metric.pkl: Loss metrics

Authors

Suk Yee Yong (sukyee.yong@csiro.au)
Julian O'Grady (julian.ogrady@csiro.au)

About

This project is being developed as part of Scientific Computing Collaboration Project 2024H1 for ERRFP 1343: Machine Learning Methods to Identify Australian Beaches from a Citizen Science derived training data.

Acknowledgements

Dataset of OpenStreetMap (OSM) beaches is created using overpass API.
Open source packages used: PyTorch, SMP, terra

Citation

If you find this repository useful, please cite the paper:

@article{yong2024sandybeachmappingdl,
  author = {Yong, Suk Yee and O'Grady, Julian and Gregory, Rebecca and Lynton, Dylan},
  doi = {10.3390/rs16183534},
  issn = {2072-4292},
  journal = {Remote Sensing},
  number = {18},
  title = {Regional-Scale Image Segmentation of Sandy Beaches in Southeastern Australia},
  url = {https://www.mdpi.com/2072-4292/16/18/3534},
  volume = {16},
  article-number = {3534},
  year = {2024},
}

License

CSIRO Open Source Software Licence Agreement (variation of the BSD / MIT License)