Whilst there are many well written articles on performing satellite image analysis, it is often a frustrating experience trying to reproduce the analysis owing to lack of detail about the package environment or broken links to imagery. This repository seeks to demonstrate best practice for allowing reproducible analysis, by using Binder to host a perfect reproduction of the analysis environment, complete with required imagery data.
Use a conda environment to isolate required packages, and write an environment.yaml file to allow Binder to recreate the environment. Conda makes installing packages such as Rasterio straightforward, handling installation of their C++ dependencies for you (GDAL!). Using Conda environments with Jupyter notebooks is a bit of a hassle, and we use https://github.com/Anaconda-Platform/nb_conda_kernels locally. Read this -> we are making an ipython kernel from our conda env (below sat_img_env) available in the base env (base) from where I normally run Jupyter.
# Setup our environment
(base) $ conda install nb_conda_kernels
(base) $ conda create --name sat_img_env python=3.6
(base) $ conda activate sat_img_env
(sat_img_env) $ conda install ipykernel
# Install dependencies
(sat_img_env) $ conda install -c conda-forge rasterio
(sat_img_env) $ conda install -c conda-forge scikit-learn
# Save our env environment
(sat_img_env) $ conda env export > environment.yml
I usually run Jupyter from my base env, and we need to make the sat_img_env env available. In a second terminal:
(base) $ python -m ipykernel install --user --name sat_img_env
(base) $ jupyter notebook
Select the sat_img_env env from the drop-down menu at Kernel > Change kernel > sat_img_env:
You are now using the sat_img_env env and can start working. EXCEPT WE ARE NOT, AND ARE STILL IN BASE.. I think this is a Mac issue, because the process works fine on Windows. anaconda/nb_conda_kernels#112
Github allows large files to be uploaded, although these should be kept to a minimum. If more than a couple of satellite images are required for analysis, it is best to put them in a public S3 bucket. I typically put imagery in a data folder.
- A good article on best practices in data science is here.