The doodleverse is an opinionated collection of Python packages designed for geoscientific image segmentation.
Package maintainers:
Currently, there are 3 packages:
- For fast, interactive segmentation of imagery
- For creation of training datasets for training Machine-Learning-based image segmentation models, for example Segmentation Gym
- A neural gym for training image segmentation models based on fully convolutional models based on UNets
- This repository is where you train models based on your own imagery and label data, and evaluate those models
- A repository of pre-trained Gym models for some common tasks
- A set of notebooks that illustrate how best to use a model on sample imagery
- You are encouraged to contribute your Gym models for the common good!
Packages are compatible, and share an underlying design and data structures
- A port of Dash-Doodler using the Anaconda/Holoviews API for graphical user interface
- Dash-Doodler and Holo-Doodler both use the same codes for image label generation, found in
doodler_engine
repo
- A mapping extension for application of Segmentation Zoo models on geospatial imagery
- Uses codes housed in s2mengine for conversion of label images (from Gym and Doodler) into geospatial formats, and for visualization of generic label images in geospatial formats in a small webGIS viewer
- A mapping extension for CoastSat using Segmentation Zoo models
- Provides a graphical browser-based environment for application of the CoastSat workflow for shoreline mapping
- Provides a graphical browser-based environment for application of Segmentation Zoo models for mapping features in satellite imagery
- Coast Train is a dataset made using Doodler, and has been used for training Zoo models that are being implemented in CoastSeg, Seg2Map, and elsewhere
- The Coast Train repo is here
- Consult the Coast Train website for more info. Data may be accessed from here
- PING-Mapper is a Python interface for reading, processing, and mapping side scan sonar data from Humminbird® sonar systems
- V2 (in development) will use Gym models to identify riverbed substrates, for the purposes of geomorphic and ecological mapping
- Check out the repo and website for more info
- collect beach grain size information using citizen scientists armed with their cell phone and any US coin
- uses a Gym model to detect the coin, so a scaling for the image can be determined for accurate grain size measurements
- a demo of that model is available
- Check out the data and website for more info
Opinionated Software is a software product that encodes a specific way of approaching a task. The Doodleverse is a well-researched end-to-end data-model pipeline for geoscientific image segmentation that has a lot of practices baked into it. In particular, the Doodleverse implements a specific data-creation and model-training pipeline. See this link and this article for more reading around this topic.