Scoring Every Street Corner: Mesh Block Livability Across Australian Capital Cities

Executive Summary

Suburb livability scores lie. They are averages that conceal enormous variation within the boundary they cover. Two homes 200 metres apart in the same suburb can have meaningfully different scores on tranquility, convenience, and lifestyle because they sit next to different things.

We built a mesh block livability scoring system to address this. Mesh blocks are the smallest geographic unit the Australian Bureau of Statistics publishes, typically covering 30 to 60 homes. We scored 145,422 residential mesh blocks across all 8 Australian capital cities.

Each block receives 8 scores: affluence, community, convenience, crime, family, hip, lifestyle, and tranquility. Each score is a weighted combination of 10 to 12 normalised input features. The inputs draw on census demographics, SEIFA socioeconomic indices, voting patterns, infrastructure data, OpenStreetMap point-of-interest counts, Google Places proximity, school quality data, and elevation.

Scores are computed within each city. They show relative livability compared to other locations in the same city. A convenience score of 8.5 in Darwin is not the same as 8.5 in Sydney. Both are in the top tier for their respective cities.

The system is live. It powers the property-level livability display on Microburbs for all capital city addresses.

The Problem with Suburb-Level Scores

Every major Australian property platform publishes suburb profiles. Walk score. School catchment rating. Safety index. Livability rank. These numbers are computed once at the suburb boundary and applied to every property inside it.

This is wrong. Suburbs are not uniform. Some are large. Many are internally diverse.

Consider Miranda in Sydney. Its mesh blocks range from 2.7 to 6.0 on tranquility. A single suburb. A spread of 3.3 points. The difference between blocks depends on proximity to the Princes Highway, distance from Westfield Miranda, and the presence of quiet cul-de-sacs versus through-roads. A suburb-level score of 5 is meaningless for a buyer choosing between two streets.

South Coogee has a standard deviation of 2.08 across its 53 mesh blocks on the affluence score. It contains both owner-occupied cliff-top homes and dense flat blocks. These are not the same demographic. Averaging them into a single number destroys the information that matters.

Woolloomooloo sits alongside Sydney Harbour. It also contains one of the largest public housing estates in New South Wales. Suburb-level scores blend both. Mesh block scores separate them.

The mesh block is a better unit. It is the smallest census geography. It was designed to contain roughly 30 to 60 dwellings and to be geographically homogenous. When we compute spatial features at the mesh block centroid, we are measuring what is actually around that cluster of homes.

But no one had done it at scale for Australian capital cities. We built this from scratch.

How We Built It

Step 1: SA1-Level Feature Pipeline

We began with SA1 (Statistical Area Level 1) geography. Australia has approximately 57,000 SA1s, each covering 200 to 800 people. The ABS publishes census demographic data at this level. It is the finest grain at which we have income, education, occupation, and housing tenure data.

For each SA1 in the 8 capital city Greater Capital City Statistical Areas, we computed 199 features across 7 categories:

Step 2: Mesh Block Spatial Feature Computation

For each of the 145,422 residential mesh blocks, we computed a separate set of spatial features at the block centroid. These features are what make mesh block scores different from simply disaggregating SA1 scores.

We used a KD-tree spatial index for efficient nearest-neighbour queries. For every centroid, we computed:

• POI counts within 500m radius: cafes, restaurants, gyms, medical, parks, green space
• Transport stops within 800m: bus, tram, train, and ferry stops counted separately
• Distance to nearest school in metres, via KD-tree on school geocodes
• Distance to nearest rail station
• Distance to nearest classified road: primary, secondary, tertiary
• Distance to nearest motorway or highway
• Proportion of OSM landuse within 500m that is residential, commercial, industrial, or green space
• Google Places category counts within 500m: cafes, nightlife, gyms, medical, childcare
• Presence of water body within 1km: ocean, harbour, river, lake
• School NAPLAN performance score for the nearest primary school

Step 3: Normalisation

Each input feature was normalised to a 0 to 10 scale before being combined into a score. We used 2nd to 98th percentile clipping. The clipping values were computed across all residential mesh blocks within each city.

Why clip at 2nd and 98th rather than min and max? Because extreme outliers distort the distribution for everyone else. The Sydney CBD has 9.1 on convenience. Without percentile clipping, the scale from min to max would compress all other blocks into a narrow band. Clipping lets the 95th percentile location still read clearly as high without the top 2% pulling the scale out of shape.

Step 4: Score Computation

Each of the 8 scores is a weighted linear combination of its input features, after normalisation. Weights were set based on face validity and sensitivity analysis. We tested multiple weight configurations on sample areas we know well: inner Sydney, inner Melbourne, and outer Brisbane. We selected weights that produced orderings consistent with local knowledge.

This is not machine learning. It is explicit, transparent, and auditable. Every weight is documented in Section 4.

The Eight Scores

Each score runs from 0 to 10. Higher is always better, with one exception noted below for crime. Below is the definition, weight structure, and top and bottom examples for each score.

Within-SA1 Variation: The Core Finding

This is the point. Every SA1 contains multiple mesh blocks. The blocks within the same SA1 share the same census and SEIFA inputs but differ on their spatial inputs. The variation in spatial inputs produces variation in scores.

The table below shows the mean and maximum within-SA1 standard deviation for each score across the full dataset of 145,422 mesh blocks.

The Castle Cove Example

Castle Cove is a suburb on Sydney's lower North Shore. It sits adjacent to Middle Harbour. The SA1 covering the southern portion of Castle Cove contains 12 residential mesh blocks. Those blocks range from 3.7 to 7.3 on tranquility. A spread of 3.6 points within a single SA1.

Why? The blocks closest to Eastern Valley Way receive road noise from a primary road and a secondary road intersection. The blocks at the end of quiet cul-de-sacs bordering the Middle Harbour bushland reserve receive almost no road noise. Both groups of homes are in the same suburb. Both groups have the same census-derived income and education profile. They are not the same place to live.

Miranda: A Suburb That Looks Uniform

Miranda in Sydney's Sutherland Shire scores a comfortable 6.8 on tranquility at the suburb level. But its mesh blocks range from 2.7 to 6.0. The blocks on the Kingsway side face a major arterial road and the noise and traffic of Westfield Miranda. The blocks on the quiet residential grid north of the centre score well above the suburb average. The suburb score conceals a 3.3-point range.

South Coogee: Affluence Variation Within a Coastal Suburb

South Coogee is widely regarded as affluent. And broadly it is. But the affluence standard deviation across its 53 mesh blocks is 2.08. That is large. The blocks on the clifftop with ocean views have census profiles consistent with high-income owner-occupiers. The blocks on the flatter inland sections, where flat blocks and higher-density rental housing sit, have profiles considerably different. Same suburb name. Different story depending on which 50 homes you are talking about.

Why Affluence and Crime Have Less Within-SA1 Variation

Affluence and crime have a mean within-SA1 standard deviation of 0.02 and 0.04 respectively. This is not a flaw. It is correct.

Census income and SEIFA data are only available at SA1 level. Every mesh block within an SA1 shares the same median income, the same SEIFA decile, the same occupation profile. We cannot fabricate block-level income data. It does not exist.

This means two blocks in the same SA1 will have nearly identical affluence scores. That is the right answer given the available data. The data shows the whole SA1 is affluent or not. It cannot tell us which individual street within the SA1 is more affluent.

We could have applied random noise to create artificial within-SA1 variation. We did not. If the data does not support a difference, we do not show one.

Data Sources

Results by City

The system covers all 8 Greater Capital City Statistical Areas as defined by the ABS. The totals below are residential mesh blocks only. Non-residential, industrial, and parkland blocks are excluded from scoring.

Score distributions differ by city. Sydney and Melbourne have the widest spreads on lifestyle and hip because they have the strongest contrast between inner urban and outer suburban areas. Darwin and Hobart have compressed distributions because the within-city variation in amenity is smaller.

This is one reason normalisation is within-city rather than national. A convenience score of 7 in Darwin describes something different from a 7 in Sydney. Within-city scoring lets a Darwin buyer understand relative convenience for their city without distortion from Sydney's density extremes.

Defence of the Approach

Why Weighted Linear Combination?

Machine learning models would produce opaque scores. A gradient boosted tree might learn that suburbs with a high ALP vote and more than 14 cafes within 500m score well on hip. That might be true in the training data. But it is not auditable. We cannot explain it to a buyer. And it will overfit.

Weighted linear combination is explicit. Every weight is published in this document. Every input is defined. If the hip score for a particular block seems wrong, we can open the formula and find out why. That auditability is worth the trade-off in predictive precision.

This is the same reason we use simpler models in the capital growth forecasting algorithm. Overfitting is the enemy of a score that needs to generalise across 145,000 locations and remain valid for several years.

Why Within-City Normalisation?

Sydney has 9.1 for the top convenience score. Darwin has roughly 7.5 for its top convenience score. These are not the same level of absolute convenience. But they are both the best their cities offer.

A buyer in Darwin is choosing between Darwin locations. They do not need to know that Darwin's top score is lower than Sydney's. They need to know which suburb in Darwin is most convenient relative to other Darwin options. Within-city normalisation gives them that.

National normalisation would compress all Darwin and Hobart scores into a narrow band at the lower end and be useless for local decision-making.

Why 2nd and 98th Percentile Clipping?

The Sydney CBD, Dangar Island, Bundeena, and Airds are real places but they are statistical outliers. Without percentile clipping, they would anchor the min and max of the distribution and compress everything in between.

After clipping at the 2nd and 98th percentile, the effective range covers 96% of all residential mesh blocks in a city. The top 2% all show as 10. The bottom 2% all show as 0. The scale is meaningful for the vast majority of locations and the extreme outliers are still correctly identified as the top or bottom of the range.

A score of 10 does not mean perfect. It means top 2% in your city.

SA1-Level Demographic Data for Block-Level Scores

This is the most common question about our approach. If all blocks in an SA1 share the same census inputs, is the block-level score really block-level?

For affluence and crime: no. Those scores are effectively SA1-level scores applied to the blocks within that SA1. We are honest about this.

For tranquility, convenience, lifestyle, and hip: yes. The dominant inputs for these scores are spatial features computed at the mesh block centroid. The census inputs play a secondary role in these scores. Two blocks in the same SA1 can and do score differently because they sit next to different roads, parks, and businesses.

For family and community: partial. School quality and community organisation proximity are block-level spatial inputs. Owner-occupier rate and income are SA1-level inputs. The mix produces moderate within-SA1 variation.

The score design reflects data availability. We use block-level inputs where they exist. We use SA1-level inputs where block-level data does not exist. We do not pretend otherwise.

Limitations

Census Data Vintage

The demographic inputs are from the 2021 census. As of March 2026, this data is 5 years old. Areas that have changed significantly since 2021, such as growth corridor suburbs in outer Melbourne and Brisbane, may have census inputs that no longer reflect current demographics. The 2026 census will replace these inputs when released.

Spatial features from OSM, Google Places, and GTFS are refreshed more frequently. The most recent updates are from 2024. These represent current conditions more accurately than the demographic inputs.

OSM Coverage Variability

OpenStreetMap coverage is excellent in major urban centres and poor in outer suburban areas. A cafe in Ultimo is almost certainly in OSM. A cafe in Yanchep on Perth's northern fringe may not be. This means convenience and hip scores in outer areas may understate reality where OSM has gaps.

We supplement OSM with Google Places API for cafes, restaurants, gyms, and medical categories. But Google Places also has coverage gaps in new developments and outer areas.

No Actual Crime Data at Block Level

The crime score uses demographic and socioeconomic proxies, not actual crime statistics. State police services publish crime data at suburb or LGA level. Some states publish at postcode. None publish at SA1 or mesh block level.

Our crime score will correlate with affluence because both use SEIFA as a primary input. A low-crime, low-affluence area will score lower on crime than it should. We are investigating whether police open data can be matched to SA1 boundaries to improve this score in a future version.

NAPLAN School Data Gaps

Not all schools publish NAPLAN results. Small schools, special schools, and some independent schools are excluded from ACARA reporting. Where NAPLAN data is unavailable, we substitute the statewide median. This introduces some noise into the family score in areas where the nearest primary school does not report results.

Static Scores, Dynamic Reality

These scores are computed at a point in time. A new motorway, a new train station, or a major development approval will change the spatial inputs for affected blocks. We plan to refresh spatial features quarterly and flag blocks that have experienced significant input changes since the last score computation.

Subjective Weight Choices

The weights in each formula were set by us. They reflect our judgement about what matters most for each dimension. Other researchers would make different choices. We consider the weights reasonable and they produce results consistent with local knowledge in our test areas. But they are not objectively correct. They are informed choices.

We publish the weights in full so they can be critiqued. We expect to revise some weights as we gather feedback from users who know specific areas well.

Generated on 2 March 2026 at 09:14:37 | Microburbs Research | Author: Luke Metcalfe