0.1.0 release and publication

[calebrob6]

The following list enumerates all tasks that need to be completed before:

• making the repo public,
• tagging a 0.1.0 release,
• publishing on PyPI/conda-forge/Spack, and
• publishing a paper on TorchGeo.

This list may change over time as we reevaluate the remaining tasks that need to be done.

Documentation

• Decide on a logo
• Add tutorial (Jupyter Notebook tutorials #38, Add tutorials to documentation #93)
• Add installation instructions (Add installation instructions #153)
• Add contributing instructions
• Document release procedure (https://github.com/microsoft/torchgeo/wiki/Releasing-Instructions)

Continuous Integration

• Set up codecov, add badge (must be public first?)
• Set up RTD, add badge (must be public first?)
• Determine minimum supported dependency versions (Determine minimum supported dependency versions #32, Document minimum supported dependency versions #80)
• Achieve 100% test coverage

Publication

• Decide on a publication venue (preprint on arXiv, paper in ICLR)
• Benchmark different ways of performing the reprojection, etc. (Benchmarking of GeoDataset for a paper result #81)
• Compare no-pretrain/imagenet-pretrain/torchgeo-pretrain
• FInd publications that discuss and compare various geospatial datasets
• Create an Overleaf document with ICLR template

Datasets

• Implement a few more GeoDatasets
  • Chesapeake (Add Chesapeake Dataset #18)
  • NAIP (Add NAIP dataset #57)
  • MSFT building footprints (Add CanadianBuildingFootprints dataset #69)
• Refactor indexing and getitem logic from datasets to GeoDataset base class (Refactor GeoDataset #73)

Models

• Set up infrastructure for hosting pre-trained model weights

Samplers

• Implement GridGeoSampler (Implement GridGeoSampler #78)
• More intelligent sampling (More intelligent sampling #82, More intelligent sampling #84)

Trainers

• Factor out task specific configuration logic into the different task classes
• Pass the "task level" arguments to the DataModules as well as the Tasks. Batch size/seed/num workers should probably go in the "task level" arguments.
• Add trainer for GeoDataset
• Refactor trainers to reduce code duplication (Reduce code duplication in trainers #205)

Transforms

• Experiment with different ways of doing augmentation pipelines
• Use Kornia, provide wrapper
• Add geospatial-specific transforms (Indices Transforms #127)