implicit-posterior

Code for Resolving Label Uncertainty with Implicit Posterior Models.

qr_for_landcover

This directory details and replicates the experimental steps for the land cover experiments.

Steps for setting up the python environment, downloading data, and running the processing and experiments for land cover mapping are detailed below. Depending on where you store the datasets, you may need to reset some of the paths in the config files in qr_for_landcover/conf (and same thing for some of the evaluation scripts and notebooks).

Setting up the environment

1. From the qr_for_landcover directory run conda env create -f environment.yml
2. Activate the environment by running conda activate qr_torchgeo
3. Update to the most recent torchgeo by running python -m pip install git+https://github.com/microsoft/torchgeo.
4. If you want to use jupyter notebooks to read and map the outputs, you'll need to configure this with the python environment.

Downloading datasets

The default parameters in this repo will assume you have data stored in /torchgeo_data. To download the datasets, you can follow these steps: Chesapeake:

• First download the ChesapeakeCVPR dataset, e.g. by following the instructions at: https://lila.science/faq#downloadtips .
• Next add the prior files by downloading and following the instrcutions from downloading https://zenodo.org/record/5866525#.YeXhCVjMIws (use version 1.1).

EnviroAtlas:

• The files for the EnviroAtlas dataset used in this analysis can be downloaded at: https://zenodo.org/record/6268150#.YhkcxpPMIws (use version 1.1).

Constructing the priors (optional):

Important: If you want to skip constructing the priors and move ahead to the experiment scripts, you can download the precomputed priors from torchgeo using the steps in the previous section. You only need to follow these steps if you explicitly want to recreate the priors from the original data.

To construct the priors for the Chesapeake dataset, first make sure you have the original dataset downloaded via torchgeo. Then, from `qr_for_landcover/compute_priors' run:

1. compute_cooccurrence_matrices_chesapeake.py to compute the class cooccurrence matrices from the training sets in each state, and then
2. make_priors_chesapeake.py to make the priors and save them in the folder from torchgeo. Note that you'll need to change the paths to the data directories at the top of each script. The notebooks in the qr_for_landcover/compute_priors will visualize these outputs.

To construct the priors for the EnviroAtlas dataset, theres a few additional steps to download the additional from the original data sources. The quick way is to download the data in the zip file from torchgeo (see above).

1. The cooccurrence matrices for the EnviroAtlas data are provided in this the data link above, so you don't have to download the full data to use them. If you're interested, the compute_cooccurrence_matrices_envirotlas.py is the script to generate them from the full EnviroAtlas data (which you'd have to download separately).
2. make_priors_envirotlas.py makes the priors and saves them in the folder from torchgeo. Note that you'll need to change the paths to the data directories at the top of each script. The notebooks in the qr_for_landcover/compute_priors will visualize these outputs.

To generate the learned EnviroAtlas priors from the inputs to the hand-coded prior:

1. Run learn_the_prior_enviroatlas.py from the experiment_scripts folder.
2. To run the model forward and save these learned priors, from the evaluation folder run save_learned_priors.py
3. You can visualize the learned priors with evaluation/visualize_output/visualize_learned_priors_ea.ipynb.

Experiment scripts:

The experiment scripts are broken up into hyperparameter search scripts (hp*.py) and evaluation runs (run*.py). To just replicate results in the paper, you can skip the hyperparameter searches. Evaluation of the results is described in the next section.

Chesapeake:

• To run the Chesapeake full experiment with hyperparameter search, run hp_gridsearch_de.py, then run_qr_in_chesapeake_north.py.

EnviroAtlas):

• To train the EnviroAtlas high-res model with just NAIP Imagery as input and with the prior concatenated as input, run hp_gridsearch_pittsburgh.py and hp_gridsearch_pittsburgh_with_prior_as_input.py
• To train the EnviroAtlas QR model with random initialization of the model weights, run hp_gridsearch_qr_from_scratch_pittsburgh.py to pick parameters in Pittsburgh and run_qr_forward_enviroatlas_from_sctrach.py to run the model in the test set in each city.
• To train the EnviroAtlas QR model with the best pittburgh model as the initializing for the model weights, run run_qr_forward_enviroatlas_from_checkpoint.py to run the model in the test set in each city.
• To use train the EnviroAtlas model using the learned prior, run run_qr_forward_enviroatlas_learned_prior_from_checkpoint.py.

Evaluating and visualizing results:

To evaluate the Chesapeake Conservancy predictions in NY and PA:

1. Evaluate the predictions against the high resolution labels with evaluation/evaluate_qr_models_chesapeake.ipynb

To evaluate the EnviroAtlas predictions in each state:

1. Evaluate the predictions against the high resolution labels with evaluation/evaluate_models_enviroatlas.ipynb

To save model output as tifs (e.g. for easy visualization):

1. Run save_predictions_chesapeake.py or save_predictions_enviroatlas.py from the evaluation folder. If you only want to evaluate some enviroatlas experiments, you'll have to comment out some lines in that script.

To visualize the outputs, use the notebooks in evaluation/visualize_output. You'll need to adjust some of the directories defined at the top of the notebook to point to where your data and output are stored.

Figures:

Notebooks to generate the figures in the paper are in the figure_notebooks folder. You'll need to adjust some of the directories defined at the top of the notebook to point to where your data and output are stored.