Speaker: Kazuki Sakamoto
Paper: arxiv.org/abs/2406.02584 // [.bib]
Transcript: Today I’ll introduce and present findings from our scoping review conducted for the forthcoming book entitled The Geography of Poverty. This work recognizes that social science research is evolving rapidly in the digital age. In recent years, we’ve witnessed a significant increase in the use of Earth observation data combined with machine learning for causal analysis and research.
Given this trend, we were motivated to explore the practices and implications of these methods for the social sciences. Despite the growing interest, this remains an emerging field without comprehensive reviews, and therefore our work fills a critical gap in understanding how these technologies are being applied.
Our review examines explicitly the types of EO data being used, prevalent methodological approaches, and their applications in causal analysis. Moreover, we consider the unique challenges that EO data presents for causal inference with careful attention to the data, methodological, and spatial dimensions since these considerations are essent essential for ensuring rigorous research outcomes. For our review methodology, we established four criteria for inclusion.
Social science relevance, inclusion of earth observation data, causal inference oriented research and machine learning methodologies. We then examined literature from 2011 to 2024, which marked significant advancements in research using both machine learning and satellite technology. Initially, we collected 142 published and preprint articles.
However, to ensure depth of analysis, we systematically narrow these down to 11 papers for in-depth examination based on our inclusion criteria. Our analysis revealed several key findings. First, considerable diversity exists in the methods applied for causal inference across studies. Second, we observe geographic diversity in particularly in regions with limited data suggesting EO data’s potential to overcome traditional data gaps. Third, we found that EO data is primarily being utilized for outcome or or coariant imputation, but there are growing publications incorporating them in other novel ways, thus indicating an untapped capacity for methodological innovation. Lastly, experimental designs became more prevalent with the expansion of randomized controlled trials.
Based on these findings, we formalized frameworks for the use of EO data in causal inference through directed as cyclic graphs where M represents a satellite imagery component. This work carefully considers the positioning of image data within causal identification strategies, thereby providing a theoretical foundation for future research. Based upon the findings from the review, we propose a three-section protocol to guide researchers when implementing EO-ML causal approaches. The first section, question and identification, helps researchers formulate their research questions to determine if they are indeed causal, identify appropriate strategies, and importantly, distinguish what the EO data is measuring.
This crucial distinction involves differentiating between direct measurements, such as deforestation or land use, and indirect, modeled measurements, like poverty or famine risk, because the nature of measurement fundamentally affects the causal interpretation. The second section, Earth observation data considerations, addresses the unique multi-dimensional nature of of Earth observation data encompassing time, width, height, and spectral bands.
Each dimension presents distinct challenges that must be carefully navigated. After specifying data requirements, researchers must consider additional factors, such as information leakage, where various elements of the causal diagram may be captured within the image. They must also address challenges related to missing data, such as those caused by cloud coverage or privacy concerns, as well as problematic image overlaps that could impact downstream analysis. Without addressing these challenges, causal analysis may be compromised from the outset.
The third section, “Causal Estimation and Uncertainty Quantification,” guides researchers in selecting appropriate computer vision models — whether spatial, temporal, or both — that can then be integrated into traditional causal inference approaches. This integration of machine learning with causal methods is particularly challenging yet essential for valid inference. Our review shows that the global potential for EO data has not yet been fully realized. Although significant progress has been made, several important considerations warrant further examination in future research. Spatial issues remain paramount, including the leakage of image-based information across spatial units. Spatial dependence captured in images, which may violate traditional causal assumptions.
Model issues present ongoing challenges, including foundation models and data biases that can propagate through analyses. Limited model interpretability conflicts with the needs of causal inference. Difficulties arise in incorporating multi-dimensional poverty metrics within machine learning frameworks.
And lastly, ethical issues cannot be overlooked, particularly the use of private and sensitive data, which raises concerns about consent, privacy, and security. In conclusion, while Earth observation data offers tremendous promise for advancing causal inference and social science research, its application requires careful consideration of methodological, technical, and ethical dimensions. Our protocol aims to provide a roadmap for researchers navigating this complex yet promising terrain. If you would like to find more information and resources, please check out PlanetaryCausalInference.org. Thank you.
References
@article{sakamoto2025scoping,
title={A Scoping Review of Earth Observation and Machine Learning for Causal Inference: Implications for the Geography of Poverty},
author={Sakamoto, Kazuki and Connor T. Jerzak and Adel Daoud},
journal={To appear in: Hall, Ola and Ibrahim Wahab (eds.), Geography of Poverty},
year={2025},
volume={},
pages={},
publisher={Edward Elgar Publishing (Cheltenham, UK)}
}–
Related Work
@article{daoud2024planetary,
title={Planetary Causal Inference: Understanding Society, Economy, and Environment through Satellite Images},
author={Daoud, Adel and Connor T. Jerzak},
journal={A Book Project},
year={2024},
volume={},
pages={},
publisher={}
}@article{fucheng2025optimizing,
title={Optimizing Multi-Scale Representations to Detect Effect Heterogeneity Using Earth Observation and Computer Vision: Applications to Two Anti-Poverty RCTs},
author={Fucheng, Warren Zhu and Connor T. Jerzak and Adel Daoud},
journal={Proceedings of the Fourth Conference on Causal Learning and Reasoning (CLeaR), Proceedings of Machine Learning Research (PMLR)},
year={2025},
volume={},
pages={},
publisher={}
}[Video][>]
@article{jerzak2023image,
title={Image-based Treatment Effect Heterogeneity},
author={Jerzak, Connor T. and Fredrik Johansson and Adel Daoud},
journal={Proceedings of the Second Conference on Causal Learning and Reasoning (CLeaR), Proceedings of Machine Learning Research (PMLR)},
year={2023},
volume={213},
pages={531-552},
publisher={}
}–