This repository provides methods for manipulations of satellite imagery, training and testing methods for object detection using detectron2 and methods for inference on satellite imagery.
Together, these aim to provide a full and open-source pipeline to infer electric infrasture on a large scale. This repository is a work in progress and only provides central features.
We are building on the dataset of fully annotated grid infrastructure in satellite imagery provided by a recent work from Duke University to train our models. However, the quality of scalable imagery can be suboptimal, so we are working on transferring that training performance to images with more scalable quality. For this, we are using approaches such as cycle-GANS and superresolution models.
As the repository joins a broad range of methods, the tools are clustered into the following directories:
- src/make_data: Methods to create annotated datasets from tif files and the respective dataframes
- src/train: Methods to train object detection networks and enhance training performance in various ways
- src/infer: Tools to check performance or conduct inference on whole .tif files
- src/utils: Helper-functions used across the other directories
Additionally, some useful workflows are provided in the notebooks directory.
Many of the methods require quite intricate environments and some even can only be executed when supported by high-performance GPUs. Therefore, we have stored dependencies for each directory separately, feel free to take a look at the directory you are interested in.
We are always happy about feedback and potential collaborators. Feel free to reach out at lukas.franken@ed.ac.uk!