Time series of satellite imagery improve deep learning estimates of neighborhood-level poverty in Africa (IJCAI 2023)

Lab | Paper | Appendix | Video | Generated maps

This is the official repository for the IJCAI 2023 paper "Time series of satellite imagery improve deep learning estimates of neighborhood-level poverty in Africa".

Authors: Markus Pettersson, Mohammad Kakooei, Julia Ortheden, Fredrik D. Johansson, Adel Daoud.

Apptainer environment

In order to improve reproducability, we ran all of our code using a single Apptainer (previously known as Singularity) container. This container can be built using the included recipe file apptainer_recipe.def as described in the apptainer documentation. Make sure you include the image path you select, e.g. path/to/image/location.sif, in your version of the configuration file config.ini.

To execute a .py script, simply run

$ apptainer run path/to/image/location.sif -nv path/to/script/file.py --script_args

in order to run one of the jupyter notebooks, you can start a jupyter lab session by running

$ apptainer exec path/to/image/location.sif -nv jupyter

Running trained single- and multi-frame models

Steps:

1. Set up your local paths and other environment variables in the config.ini file.

2. Download the satellite data, calculate the dataset variables and prepare the cross-validation folds as outlined in the preprocessing directory.

3. Make predictions for the different pretrained models by running inference_model.py. In case your system is equipped with Slurm, you can simply run the inference_model.sh script

4. Generate the figures as presented in the paper by running the evaluate_results/model_evaluation.ipynb and evaluate_results/ts_effect.ipynb notebooks.

Acknowledgements

Preprocessing and evaluation code in this repository takes a lot of inspiration from the work by Yeh et al., creators of the architecture we call "single-frame model". You can find their codebase here.

Citation

Please cite our paper as

Markus B. Pettersson, Mohammad Kakooei, Julia Ortheden, Fredrik D. Johansson, & Adel Daoud (2023). Time Series of Satellite Imagery Improve Deep Learning Estimates of Neighborhood-Level Poverty in Africa. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23 (pp. 6165–6173).

Or use the follwoing BibTex entry

@inproceedings{pettersson2023time,
  author       = {Markus B. Pettersson and
                  Mohammad Kakooei and
                  Julia Ortheden and
                  Fredrik D. Johansson and
                  Adel Daoud},
  title        = {Time Series of Satellite Imagery Improve Deep Learning Estimates of
                  Neighborhood-Level Poverty in Africa},
  booktitle    = {Proceedings of the Thirty-Second International Joint Conference on
                  Artificial Intelligence, {IJCAI-23}},
  pages        = {6165--6173},
  publisher    = {International Joint Conferences on Artificial Intelligence Organization},
  year         = {2023},
  month        = {8}
  url          = {https://doi.org/10.24963/ijcai.2023/684},
  doi          = {10.24963/ijcai.2023/684}
}