dipan saha

copywriter, liberal, futurist, skeptic

[dipan saha @s_dipan]

EO is a technology that allows planetary vision. Sovereignty stifles that vision from flourishing.

[Aravind 🌍 🛰 @aravindEO]

Sovereignty is perhaps of the one of the more paradoxical ideas in Earth observation today:

One, because it is an industry that basically runs on an open-source software stack - most EO products or EO-derived solutions uses/runs on Python, R, GDAL, Xarray, PyTorch, Zarr, GeoPandas, Leaflet, QGIS, Pangeo, STAC etc.

These libraries and tools are built and maintained by a truly global community - one can 'sovereignize' where the infrastructure runs, but not the community that is building and continuously improving the technology underneath it.

Two, because a truly global EO solution requires global cooperation and collaboration among public, private and research community to advance across the value chain - from standards for data acquisition, management and processing to calibration/validation and training data collected across geographies and shared through common global frameworks.

I know that we are going through an unprecedented time geopolitically and it requires every government, company and organisation to look inward. But the mistake, from an EO perspective, is to confuse sovereignty (control) with autarky (self-sufficiency).

EO is global by nature, and some of the infrastructure that makes sovereign EO capabilities possible is itself built through planetary-scale systems and communities that are transcend national boundaries.

[dipan saha @s_dipan]

@AmrullahSaleh2 For India, this could be a concern. Isn't it?

[Aravind 🌍 🛰 @aravindEO]

https://t.co/7GwfqfPSwO

[Erin Burnett OutFront @OutFrontCNN]

A CNN investigation reveals Russia likely provided crucial satellite information about U.S. bases, giving Tehran a major advantage in the war.

[Yohan @yohaniddawela]

Google turned 46,000 places into 330-number fingerprints, then used them to predict where people live better than maps of nightlights, roads and buildings.

That’s a strange sentence, but it points to a pretty big shift in population mapping.

For decades, the standard approach has been to estimate population from things we can see from above: night-time lights, road networks, buildings, land cover, accessibility, elevation, slope, and other physical traces of settlement.

The basic idea is that people tend to live where there are houses, roads, lights, services, and accessible land.

But the visible world can actually mislead us.

A bright place can be an industrial zone. A dense road network can serve commuters rather than residents. Two built-up areas can look similar from space while having completely different household sizes, incomes, migration patterns, or settlement histories.

A new paper from
@WorldPopProject
tests Google’s Population Dynamics Foundation Model.

Instead of feeding the model a fixed list of hand-built geography variables, PDFM gives each place a 330-dimensional embedding. You can think of an embedding as a compressed fingerprint: a set of numbers that captures patterns too complex to write down manually.

Those 330 numbers come from several sources.

The first 128 dimensions come from aggregated search trends. The next 128 come from maps and busyness indicators. The final 74 come from weather and air-quality signals.

Then the model links places together in a graph, where districts, postal codes, ZIP areas and local government areas are treated as connected nodes. Places can be connected because they’re geographically close, because they sit inside the same administrative structure, or because their behavioural signals look similar.

That means the model is actually learning the wider pattern of how that place functions.

The authors tested this across three very different countries:

• Nigeria: 783 local government areas
• Brazil: 5,568 districts
• United States: 39,649 ZIP Code Tabulation Areas

Together, that’s 46,000 places.

The benchmark was a standard WorldPop-style feature set: 23 geospatial covariates covering the built environment, night-time lights, accessibility, land use, land cover, elevation, slope and related variables.

The models themselves were deliberately ordinary: Random Forest, XGBoost and Elastic Net. Interestingly, the paper wasn’t trying to win by hiding everything inside a more complex model. It was testing whether the predictors themselves carried useful information.

The evaluation was also stricter than a simple random split.

The authors used spatial cross-validation, grouping smaller places by larger administrative regions. Nigerian LGAs were grouped by state, Brazilian districts by state, and US ZIP areas by county.

That’s important because neighbouring places often share the same roads, markets, lights, infrastructure and settlement patterns. If a model trains on one area and predicts the area next door, it can look more accurate than it really is.

Spatial cross-validation makes the model prove that it can travel.

Across the analyses, the PDFM embeddings consistently improved predictive performance relative to the standard geospatial covariates.

In plain English: the learned place 'fingerprints' contained population-relevant information that the usual satellite and accessibility layers were missing.

That’s a pretty interesting result because population maps sit underneath a lot of real-world decisions.

They shape health planning, disaster response, poverty measurement, climate exposure analysis, infrastructure investment and humanitarian targeting. A small spatial error can change where clinics are planned, where vaccines are sent, where flood risk is counted, and where public money goes.

A pretty interesting part of the paper is its restraint.

The authors thankfully don’t claim that foundation model embeddings make old geospatial features obsolete. Their finding are a bit more useful than that.

PDFM helped most when common physical indicators of settlement intensity and accessibility didn’t already explain the population pattern well. In places where night-time lights, buildings, roads and accessibility were already highly informative, the traditional covariates remained competitive.

That’s exactly the kind of result a good population model should produce.

Buildings matter. Roads matter. Lights matter. Search behaviour, busyness patterns, weather context and environmental signals can matter too.

The real advance is learning when each signal adds value.

This is where the paper becomes bigger than one benchmark. Population modelling is moving from a world built around visible settlement proxies to one built around richer place representations.

The old question was: what can we see from space?

The new question is: what combination of physical, behavioural, environmental and administrative signals best explains where people actually live?

That shift matters because the census is still the anchor, but censuses are expensive, uneven, delayed and sometimes politically difficult. Satellite layers help fill the gaps, but they mostly capture the visible surface of settlement. Embeddings add another layer: the learned context of human activity around a place.

The future probably belongs to hybrid models.

Use the census where it exists. Use satellite covariates where the physical landscape is informative. Use learned embeddings where behaviour, accessibility and local context reveal things the visible landscape leaves out.

I find that to be a more realistic path than replacing one data source with another.

Link to paper: https://t.co/Z0YefvJEzS

[Yohan @yohaniddawela]

Rice cultivation in Asia consumes 30% of the world’s exploited freshwater and is a massive source of methane emissions.

But until now, we lacked granular data to track it.

A AI model from NASA and IBM has finally mapped the entire continent at 30 m resolution

[Andrew Hessel @andrewhessel]

A cell is like a factory. It can be programmed to produce many different and useful products, such as insulin or antibodies. And it can scale by growing and dividing. The cell's genome is the operating system—the full set of instructions for building and operating the cell.

[West Bengal Radio Club (VU2WB) @_vu2wb_]

#pmo #indianrailway #suvenduadhikari

[dipan saha @s_dipan]

Why it can't be built in West Bengal? The project just requires 1.4 km water front.

https://t.co/KIdQSuzYaU

@SuvenduWB

[dipan saha @s_dipan]

@PeterDiamandis That's the beginning of an end to whatever that makes America greate. Europeanisation of USA.

[dipan saha @s_dipan]

@gareebscientist @gareebscientist any idea about expected commercial cost of vikram-2 of a single launch and total orbital insertion capacity?

[dipan saha @s_dipan]

@gareebscientist @INDGeopolitics @VishnuNDTV @gareebscientist problem is not talent in India. Cuttingbedge innovation requires "fail fast and fail more" attitude in management. India has a resource problem. We don't appreciate failures.

[dipan saha @s_dipan]

@ajay_shah Sir, sorry for an unrelated query. Can you suggest a decision matrix to choose which city to migrate to within India for better quality modern edu, entrepreneurship and career scope? Pop choice is Bangalore. But what about Pune or Mangalore or others?

[Ajay Shah @ajay_shah]

https://t.co/HLhUTF3vBV

[dipan saha @s_dipan]

@gareebscientist Very appropriate.

[dipan saha @s_dipan]

@rajbhagatt Great. I and my wife have business interests in the same field. So, we might disturb you in future. Hope you won't mind.

[dipan saha @s_dipan]

@gareebscientist Oho. Really bad news. I keep more faith on Bezo than Musk.

[Ajay Shah @ajay_shah]

https://t.co/ftaQ7PwBNZ

[dipan saha @s_dipan]

@PixxelSpace @pintofviewclub Will it be uploaded on YouTube later?