Planetary Causal Inference: Why Now?
Planetary causal inference is emerging as a field of inquiry focused on the integration of planetary-scale Earth observation (EO) data with localized RCTs and natural experimental studies---why now? Researchers and companies have been running RCTs in large number since the mid-1950s; governments have incorporated satellite imagery into decision-making since around the same time---the Landsat program is nearly half a century old.
Yet, the difference with now is that the convergence of data availability and modeling capabilities---simultaneous expansions in both the size and quality of datasets and in the modeling tools at our disposal---has opened up possibilities for the broader inclusion of Earth observation data into the social, environmental, and economic sciences.
Indeed, recent expansions in sensing data from Sentinel, or from the commercial side: Maxar with its extremely high-resolution images and Planet Labs with its near-daily coverage for almost the entire globe. Where we once had a handful of images per location with poor resolution, we can call on new data every day, with active sensing technologies now able to see through cloud cover. And so for every place in the world, there is no shortage of image coverage for discerning what is happening and when.
But data alone do not suffice. The other half of the story is in models: it is now practical, on consumer-grade hardware, to train or fine-tune medium-sized computer vision models on relevant tasks. More than that, the training methods themselves have advanced.
There are now numerous domain-specific or general-purpose foundation models for Earth observation that are trained on masked autoencoding objectives. In these models, neural networks form powerful representations of each image, forming these representations by forcing the network to learn how to predict masked portions of an image from the unmasked portions. The result is that we can embed images in a lower-dimensional representation space that preserves the relevant content for many downstream tasks. We can detect roads, building footprints, vegetation cycles, or more ephemeral features. It is precisely these subtle patterns that can prove critical for capturing effect heterogeneity in interventions or characterizing confounding factors. [1]
Effect heterogeneity is one demonstration of the value of a Planetary Causal Inference perspective. Whether for a local anti-poverty program or for new environmental regulations, we often want to know how the treatment’s effect differs across neighborhoods or regions: “average” effects might obscure important variation. But effect heterogeneity, as a matter of design, needs an input that is stable, consistent, and hopefully standardized, so it can generalize beyond the local context. Earth observation is precisely that. A multi-year satellite record has no notion of municipal boundaries; it passively collects. It doesn't stop collecting... Freed from concerns about patchy local data or one-off surveys, we can examine policy effects in a broader, harmonized view. [1] [2]
Similarly, new developments in planetary-scale data help with confounder adjustment. In observational or quasi-experimental settings, one big worry is correlation confounded with unobserved context: roads cause development---but perhaps the same factor that leads governments to build roads also fosters better schooling or health, and so on. If an unobserved variable influences both the selection of who receives an intervention and the outcome of interest, we get the usual correlation-does-not-imply-causation fiasco. But if those unobserved factors have visible signatures—location, building density, farmland expansion, sprawl, even the color or texture of rooftops—then an Earth observation time series can partially capture them. Mountainous farmland is not quite the same as coastal farmland; your unobserved confounder might now be at least partly observed in EO data through the lens of a properly trained model. [1]
In short, if the question is “Why now?” the short answer is that both the data axis and the modeling axis have greatly expanded for PCI. Earth observation data have become more frequent, more detailed, cheaper, and more systematically archived (Sentinel, Landsat, Planet Labs, Maxar, etc.). The rapid progress in large-scale deep learning has simplified the discovery of predictive relationships from unstructured images and allowed us to embed them in representation spaces that standard causal inference methods can incorporate.
Even so, obstacles remain. Some interventions might not leave strong footprints in imagery if the relevant confounders or local conditions are invisible. Clouds, atmospheric haze, or ephemeral data gaps can impede coverage. Overlapping images can cause correlated residuals. And, general transportability analysis is still an assumption-laden enterprise: if the new region’s social or political features differ in intangible ways not captured by the images, your estimates may still break down. But on the margin, these new methods can help generate new, contextually informed causal estimates in challenging conditions.