Skip to content

Doodleverse/segmentation_zoo

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

85 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

📦 Segmentation Zoo 🐘

Last Commit Maintenance Wiki GitHub Wiki Python TensorFlow Keras

zoo

Hi! This is work in progress - please check back later or use our Discussions tab if you're interested in this project and would like to say hi. The models here should be considered beta, and could improve over time, or they may change in structure, so any usage of these models should be for practice with Segmentation Gym. Better documentation is forthcoming.

🌟 Highlights

  • Segmentation Zoo is a repository of image segmentation models for broad use in the geosciences, pre-trained using Segmentation Gym. These models may be used for general scientific enquiries or for use in downstream applications (credit is always appreciated!)
  • 'Broad use' is open to interpretation, but perhaps relates to 'broad classes' of widespread utility in geosciences (including geology, ecology, hydrology, oceanography and all related fields) such as water, sediment, vegetation, soil, rock, and landcover types. Such broad classes may also be broken down into other generally useful classes, such as water types (e.g. whitewater, turbid water), sediment types (sand, gravel, etc), vegetation types, soil types, and rock types. Examples of broad, established ontologies for geoscience disciplines are CMECS and GLanCE
  • Segmentation Zoo promotes use of models published using Zenodo, along with a model card in a specific format that documents the model, its construction, and intendd uses.
  • Finally, 'Zoo contains various examples of model implementations. 'Implementation' in this scope refers to the use of a model on unseen sample imagery. There are a number of ways that this may be acheived, and would differ depending on factors such as the type of imagery, amount of overlap in imagery, required accuracy, the need for probabilistic outputs, number of classes, etc. Whereas 'Segmentation Gym' contains a basic model implementation (called seg_images_in_folder.py), bespoke model applications may differ considerably. The notebooks and scripts folders in Zoo contain basic and more advanced examples of how to use a model for segmentation, including use of ensemble models. We also hope to demonstrate possible transfer learning scenarios

✍️ Authors

Package maintainers:

Contributions:

We welcome collaboration! Please use our Discussions tab if you're interested in this project. We welcome user-contributed models trained using Segmentation Gym, but please wait for announcements while we get our act together, and meanwhile we'll be happy to chat about your data, models, science, art, and life in general in Discussions.

⬇️ Installation

If you already have the conda environment, gym, installed from Segmentation Gym, you may use that. Otherwise, we advise creating a new conda environment to run the program.

  1. Clone the repo:
git clone --depth 1 https://github.com/Doodleverse/segmentation_zoo.git

(--depth 1 means "give me only the present code, not the whole history of git commits" - this saves disk space, and time)

  1. Create a conda environment called zoo
conda create -n zoo python=3.9
conda activate zoo
conda install -c conda-forge "tqdm>=4.64.1" scipy numpy scikit-image jupyterlab joblib pandas plotly natsort matplotlib -y
pip install tensorflow doodleverse_utils==0.0.29 ipyfilechooser aiohttp
pip install tensorflow-gpu

If you get errors associated with loading the model weights you may need to:

pip install "h5py==2.10.0" --force-reinstall

and just ignore any errors.

Segmentation Zoo User Roles

  1. User of models
    • We anticipate that most Zoo users will simply want to use published models listed in the wiki, and adapt the example notebook workflows for their own purposes
  2. Contributor of models
    • We welcome contributions of Gym models! The basic steps (to be outlined in the wiki) involve a) making a new Zenodo release including your model wrights file, config file, and modelcard file; b) cloning the Zoo wiki and adding a page explaining the model's purpose; and c) issuing a pull request for review by Doodleverse HQ
  3. Contributor of implementation workflows
    • We welcome contributions of Gym model implementations! The basic steps (to be outlined in the wiki) involve a) cloning the Zoo repo and adding a notebook or script showing your workflow; and b) issuing a pull request for review by Doodleverse HQ

Example python script

To run the example script, change directory (cd) to the scripts directory and run the python script using

python select_model_and_batch_process_folder.py

You select a task, then a model, then select a folder of images to apply it to. It will write output files into a subfolder of the input imagery called 'out'

The set of available tasks.models are listed below, by theme and more details may be found on the Models pages of the wiki for satellite, aerial, and orthomosaic imagery

Task: satellite-derived shoreline location extraction

  • 2-class models
    • "sat_RGB_2class_resunet_7865364", 2 class (water, other), RGB satellite imagery
    • "sat_5band_2class_7448390", 2 class (water, other), RGB+NIR+SWIR satellite imagery
    • "sat_NDWI_2class_7557072", 2 class (water, other), NDWI satellite imagery
    • "sat_MNDWI_2class_7557080", 2 class (water, other), MNDWI satellite imagery
  • 4-class models
    • "sat_RGB_4class_segformer_7933015", 4 class (water, whitewater, sediment, other), RGB satellite imagery, segformer
    • "sat_RGB_4class_resunet_6950472", 4 class (water, whitewater, sediment, other), RGB satellite imagery, resunet
    • "sat_5band_4class_7344606", 4 class (water, whitewater, sediment, other), RGB+NIR+SWIR satellite imagery
    • "sat_NDWI_4class_7352859", 4 class (water, whitewater, sediment, other), NDWI satellite imagery
    • "sat_MNDWI_4class_7352850", 4 class (water, whitewater, sediment, other), MNDWI satellite imagery
    • "sat_7band_4class_7358284", 4 class (water, whitewater, sediment, other), RGB+NIR+SWIR+NDWI+MNDWI satellite imagery

Task: masking water in aerial images

  • "pcmsc_wm_v2_7700430", 2 class (water, other), RGB aerial oblique imagery, 1024x768 pixel imagery, westcoast high-relief coasts
  • "aerial_2class_6234122", 2 class (water, other), RGB aerial oblique imagery, 1024x768 pixel imagery
  • "aerial_2class_6235090", 2 class (water, other), RGB aerial oblique imagery, 1024x768 pixel imagery
  • "noaa_2class_7604083", 2 class (water, other), RGB aerial oblique imagery, 1024x768 pixel NOAA ERI imagery

Task: generic classification of aerial images

  • NOAA models
    • "noaa_4class_7631354", 4 class (water, sediment, vegetation, developed), RGB aerial oblique imagery, 1024x768 pixel imagery
  • FloodNet models
    • "floodnet_10class_7566810", 768x512 pixel imagery, 10 class (Background, Building-flooded, Building-non-flooded, Road-flooded, Road-non-flooded, Water, Tree, Vehicle, Pool, Grass), RGB UAV imagery
    • "floodnet_10class_7566810", 1024x768 pixel imagery, 10 class (Background, Building-flooded, Building-non-flooded, Road-flooded, Road-non-flooded, Water, Tree, Vehicle, Pool, Grass), RGB UAV imagery
  • Elwha River alluvial models
    • "elwha_alluvial_driftwood_segformer_7933013", 2 class (other, wood), RGB aerial orthomosaic imagery, 768x768 pixel imagery, segformer
    • "elwha_alluvial_driftwood_resunet_8072293", 2 class (other, wood), RGB aerial orthomosaic imagery, 768x768 pixel imagery, resunet

Task: coastal landcover classification in Orthomosaic / NAIP (Seg2Map)

  • Coast Train models

    • "ortho_2class_7574784", 2 class (water, other), 768x768 pixel RGB NAIP/orthomosaic imagery
    • "ortho_5class_7566992", 5 class (water, whitewater, sediment, bare terrain, other terrain), RGB NAIP/orthomosaic imagery
    • "orthoCT_8class_7570583", 8 class (water, whitewater, sediment, bare terrain, marsh veg, terrestrial veg, ag., dev.), 768x768 pixel RGB NAIP/orthomosaic imagery
  • Chesapeake models

    • "chesapeake_7class_7576904", 7 class (water, tree canopy / forest, low vegetation / field, barren land, impervious (other), impervious (road), no data), 512x512 pixel RGB NAIP/orthomosaic imagery
  • Barrier islands models

    • Barrier islands: substrates (forthcoming)
    • Barrier islands: geomorphic category (forthcoming)
    • Barrier islands: vegetation type (forthcoming)
    • Barrier islands: vegetation density (forthcoming)
    • Barrier islands: general classification (forthcoming)

Task: generic landcover classification (Seg2Map)

  • OpenEarthMap models
    • "openearthmap_9class_7576894", 512x512 pixel imagery, 9 class (bareland, rangeland, dev., road, tree, water, ag., building, nodata)
  • DeepGlobe models
    • "deepglobe_7class_7576898", 512x512 pixel imagery, 7 class (urban, ag., rangeland, forest, water, bare, unknown)
  • EnviroAtlas models
    • "enviroatlas_6class_7576909" 1024x1024 pixel imagery, 6 class (water, impervious, barren, trees, herbaceous, shrubland)
  • AAAI/building models
    • "aaai_building_7607895" 1024x1024 pixel imagery, 2 class (building, other)
    • "aaai_floodbuilding_7622733" 1024x1024 pixel imagery, 2 class (flooded building, other)
  • XBD/building models
    • "xbd_building_7613212" 768x768 pixel imagery, 2 class (building, other)
    • "xbd_damagedbuilding_7613175" 768x768 pixel imagery, 4 class (building/no-damage, building/minor-damage, building/major-damage, other)

Example notebook

To run the example notebooks, change directory (cd) to the notebooks directory and launch jupyter using

jupyter notebook

The menu of notebooks can be accessed in the browser at http://localhost:8888/tree

You should adapt the workflows shown in these notebooks to your own imagery

User guide

Check out our wiki for further instructions

About

A collection of geoscientific image segmentation models

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published