Scene-to-Patch Earth Observation: Multiple Instance Learning for Land Cover Classification

This repo contains the code for the paper Scene-to-Patch Earth Observation: Multiple Instance Learning for Land Cover Classification, published as part of the Tackling Climate Change with Machine Learning Workshop at NeurIPS 2022.

We use Bonfire for the backend MIL functionality.

Below we break down each of the directories in this repo:

Config

Configuration for model parameters.

Models

Contains the trained model files. Five repeats per configuration.

Out

Interpretability outputs and other figures that are used in the paper

Results

Raw results for our experiments: scene-level RSME and MAE, patch-level mIoU, and pixel-level mIoU.

Scripts

Contains our executable scripts. These are the entry points to our experiments, and should be run from the root of the repo.

Src

Contains our project-specific code. This includes the dataset, high-level model implementations (low-level model code is implemented in Bonfire), and interpretability studies.