Urban Data Sources

Start with the raw materials of data-driven urbanism, because each one has a native strength you should use and a native blindness you must remember. GIS and the cadastre are the official spatial record - plots, roads, land-use zones, ownership, infrastructure - held by the planning authority or revenue department. It is the backbone of any urban model: authoritative, structured, legally consequential. But it records the *formal* city, the one that has been surveyed and titled, and it is often out of date and patchy exactly where change is fastest.

OpenStreetMap is a crowd-sourced world map - streets, buildings, points of interest, mapped by volunteers. It is astonishingly rich, free, and in some places better than any official source; it is also uneven, because coverage follows the mappers. Well-mapped neighbourhoods tend to be wealthier, more central, more visited. A quiet informal settlement may be a near-blank on OSM not because nothing is there but because no one mapped it.

Census data counts people and households - population, age, occupation, amenities - at coarse spatial units, on a slow cycle (India's decennial census). It is the bedrock of who lives where, but it is a snapshot years old by the time you use it, it aggregates to protect privacy so fine grain is lost, and it systematically misses the mobile, the homeless and residents of settlements not on the enumeration schedule.

Movement and mobility data - mobile-phone location, GPS traces, transit-card taps, ride-hailing logs - reveal how the city actually moves, at a temporal resolution no census can match. They are among the most powerful new sources and among the most biased: they see the people who carry smartphones, hold bank-linked cards and appear in commercial datasets, and under-see everyone else. Remote sensing (satellite and aerial imagery) reveals built form, growth, green cover and heat from above, cheaply and everywhere - but it sees roofs, not lives, and cannot tell a thriving lane from a slum earmarked for clearance. Open and civic data - budgets, permits, complaints, sensor feeds - can be gold, where it exists, is published, and is clean enough to use.

Now the honest core of the lesson. Line the sources up and a pattern appears: they all tend to miss the same place, and it is the place that houses the most people. The informal city - the dense old fabric, the unauthorised colonies, the vast settlements that shelter hundreds of millions across India and the global South - falls into the gap in almost every dataset at once. It has no clean cadastre because it was never formally surveyed or titled. It is thin on OpenStreetMap because fewer volunteers map it. It is undercounted in the census because enumeration schedules and addresses do not reach it well. It is under-seen in mobility data because its residents are less likely to appear in the commercial and smartphone streams those datasets draw on. From a satellite it may read as an undifferentiated grey mass, legible as 'to be cleared' but not as a living community.

This is not a random error you can average away; it is a systematic bias that runs the same direction every time, and its direction is against the poor. A model trained or built on this evidence base inherits the blindness. A generative masterplan optimizing 'the city' is really optimizing the city its data can see - and the fabric it cannot see is precisely the fabric most at risk of being planned over, because in the model it barely exists. The danger is not merely inaccuracy; it is that invisibility in the data becomes erasure on the ground, dressed as objectivity.

The discipline this demands is simple to state and hard to practise: for every source, ask *who and what is missing, and which way does the gap run?* Treat data as a partial, made record with an author and an agenda, not as the city itself. Where a source is silent, that silence is information - usually about who did not get counted. And when the evidence base cannot see a community, the humane response is to go and look, to bring in local and participatory knowledge, and to refuse to let the map's blank spaces license decisions about the people living in them.

Bias in urban data is rarely a single villain; it accumulates quietly along a chain, and by the end it is invisible. Follow a mobility dataset from signal to claim. The raw signal is a stream of phone pings - already biased, because it only exists for people with phones and data plans. Cleaning drops points flagged as errors, snaps traces to known roads and fills gaps with assumptions; each choice is reasonable and each bakes in a worldview about what a 'normal' trip looks like. The cleaned data becomes a metric - say, average commute time - which compresses a messy human reality into one number. That number becomes a claim: 'the city's commute is twenty minutes', quoted in a plan as fact. By the last step the phones that were never in the sample, the trips snapped onto the wrong road, and the assumptions in the fill are all gone from view. The number looks objective precisely because its manufacture is hidden.

Three biases recur and are worth naming. Coverage bias: who is in the dataset at all (smartphone owners, titled property, mapped streets) skews toward the wealthier and more formal. Categorisation bias: the messy real world is forced into tidy boxes, and whatever does not fit a box - a shop-house that is home, workshop and store at once, a street that is also a market and a living room - is mangled or dropped. Temporal bias: data ages, and a census or land-use map years old describes a city that has already moved on, especially where change is fastest.

The practical response is not to reject data - data-driven urbanism is genuinely powerful and here to stay - but to read it like a critic. Trace every confident number back up the chain to the signal it came from. Ask what was sampled, what was dropped, and what was assumed. Cross-check sources against each other and against your own feet on the ground. Report uncertainty and gaps as first-class results, not footnotes. Data can sharpen judgement wonderfully; it cannot replace it, and the moment a number stops being questioned is the moment it starts to mislead.

Put this to work as a practical method. A trustworthy urban evidence base is not the biggest pile of data; it is a triangulated one, where independent sources check each other and their gaps are mapped as carefully as their contents. Combine the official cadastre (formal structure) with OpenStreetMap (fine-grained texture), census (who lives here), mobility (how they move), remote sensing (built form and change) and civic data (how the city is run and complained about) - and treat the places where they disagree as the most interesting places, not errors to smooth over. A settlement that is dense on satellite imagery, blank on the cadastre and thin on OSM is telling you exactly where the formal record fails and the people are.

Add the source no dataset contains: ground truth and local knowledge. Walk the area. Talk to residents, local officials and community organisations. Bring participatory mapping in, so the people who live in the data's blind spots can put themselves on the map. This is not softness; it is rigour - it is how you catch the systematic errors that no amount of computation will surface, because the computer only knows what it was given. In the Indian context, where informality is the norm rather than the exception across much of the urban fabric, this ground-truthing is not optional; it is the difference between an evidence base that serves the whole city and one that quietly designs for half of it.

Finally, keep the boundary from the course's first lesson. Data informs; it does not decide. An evidence base helps you understand a place, argue from something firmer than opinion, and open options for public debate - but the binding decisions about land use, displacement and a city's future belong to the planning authority, the statutory and participatory process, the affected communities and the governing law and development-control regulations, in India the master-plan process, the applicable DCR and the National Building Code of India. The most honest thing a good dataset can do is show you clearly, and in public, both what it knows and what it cannot see.