01GIS & Network Analysis2026

Modeling Residential Waste Transportation in New York City

The city reports how much waste each district produces. It does not publish how that waste moves. I built three routing models over the same street network to test what different assumptions produce.

Type: Academic research
Advanced GIS, Barnard College
Role: Co-author. Data pipeline, routing models, spatial statistics, cartography
Collaborator: Nina Feldman
Tools: ArcGIS Pro · Network Analyst · Python
Statistics: Local Moran’s I · Lee’s L

Dense web of green candidate waste-truck routes across all five boroughs of New York City, converging on black disposal site markers — Fig. 01 **The routing universe.** Five closest-facility candidate routes from every district section. All three models select from this fixed set.

Question: How do different routing assumptions change the modeled geography of waste transport, and where do the resulting corridors fall?
Approach: One fixed truck network and one candidate route set. Three assignment rules. Compared with Local Moran’s I and Lee’s L.
Result: The three rules produce materially different maps, but four corridors persist across all of them. The most flexible rule clustered the hardest.

Premise

Sanitation is one of the most visible infrastructural systems in New York and one of the least documented at route level. You can find out how many tons a district produced last month. You cannot find out how those tons travelled.

That absence has analytical consequences. Represented only as an aggregate quantity per district, the spatial cost of moving waste disappears. Route intensity, corridor exposure, and any relationship between transport patterns and demographic conditions become impossible to examine.

The Department of Sanitation’s operational routes are not public, so this project does not try to reconstruct them. It models three plausible routing logics and compares them. The comparison is the experiment.

What do different routing assumptions produce when you follow each one all the way through to a map?

Research question

Data preparation

DSNY reports refuse tonnage by district. Districts are too coarse to act as collection origins, so tonnage has to be distributed to district sections, the finer operational unit.

I aggregated residential unit counts from MapPLUTO at both scales, then allocated each district’s annual tonnage across its sections in proportion to residential share. Section centroids became the modeled route origins.

First assumption: Residential units stand in for waste generation, which assumes relative uniformity within a district.
Known distortion: Filling smaller units from a larger aggregate introduces the Modifiable Areal Unit Problem. Every downstream number is a modeled approximation.

Four small maps: DSNY districts, DSNY sections, a circular inset of PLUTO residential building footprints, and a blue choropleth of residential density by section

Choropleth map of New York City showing estimated waste tonnage allocated to each DSNY district section in four shades of green

Network construction

I built a truck-eligible road network from NYC LION street centerlines in ArcGIS Pro, then ran a closest-facility analysis from every section centroid to its five nearest disposal or transfer facilities.

That output is the routing universe, the constrained set of possibilities all three methodologies draw from. Fixing it first is what makes the later comparison fair: the three models differ only in how they select from this set.

Staten Island is the clearest constraint in the network. A sparse loop tied to a single facility, with one long connection over the Verrazzano that no routing logic can avoid.

Three small maps in a row: graduated section centroids as origin points, black disposal facility destinations, and the pale grey truck-eligible LION street network — Fig. 04 **Three inputs.** Origins graduated by estimated tonnage, in-city disposal and transfer destinations, and the truck-eligible street network.

Three routing models

Each model resolves the same question differently. Method 1 optimizes for proximity alone. Method 2 introduces facility capacity through a Python optimization routine. Method 3 drops single-facility assignment and splits each section’s tonnage across two facilities.

Routes were then buffered 25 ft, clipped to census tracts, and weighted by tonnage and length within each tract. That route intensity surface is the unit of every analysis that follows.

Legend: route tonnage graduated from small to extra large, district section centroids, and disposal sites

Method 1 route map: green routes radiating outward around many disposal facilities across New York City

Method 2 route map: fewer, heavier trunk corridors converging on larger-capacity disposal facilities

Method 3 route map: many overlapping route corridors from tonnage split proportionally between two facilities

Spatial clustering

Three Local Moran's I cluster maps of New York City on black backgrounds, comparing route intensity clusters and outliers across the three routing methods — Fig. 05 **Clusters and outliers by method.** High-intensity clusters hold in the same corridors under all three rules while the texture of clustering shifts.

Four areas surface under every model: the South Bronx, Upper Manhattan, Staten Island, and parts of North Brooklyn. Because they persist across three different assignment rules, they point to facility siting and network topology rather than to any one routing logic.

The differences are equally informative. Method 1 fragments intensity radially around many destinations. Method 2 consolidates it near larger facilities. Method 3 distributes waste across more facilities and still produces the strongest clustering of the three.

Splitting tonnage across more facilities produced the strongest corridor concentration of the three.

Finding · Method 3

Demographic association

Lee's L bivariate maps comparing Supplemental Security Income percentage against route intensity for each of the three routing methods — Fig. 06 **Supplemental Security Income against route intensity.** The greyscale panel is the demographic surface. The three colored panels are its spatial association with route intensity under each method.

One result has to be read carefully. Total population correlates with route intensity partly by construction, since residential units were the proxy used to allocate tonnage in the first place. Reporting that as a finding would be circular, so I set it aside.

The Supplemental Security Income surface is the interpretively useful one. Under all three methods, modeled route intensity co-occurs with areas of socioeconomic vulnerability, particularly in the South Bronx, North Brooklyn, and Southeast Queens.

This identifies a spatial coincidence in a modeled surface. It does not establish exposure, health burden, or causation.

Limitations

When the operational system is hidden, the model is a structured hypothesis. These are its terms.

Proxy allocation. Section tonnage is estimated from residential units. MAUP uncertainty carries through everything downstream.
Centroids are not pickup points. They compress the distributed reality of curbside collection into one origin per section.
Static shortest paths. The models ignore truck capacity, traffic, time-of-day routing, sanitation schedules, marine transfer timing, garage locations, and collection sequencing.
First-degree destinations only. Many facilities modeled here are transfer stations rather than final disposal. Waste geography beyond city limits is out of scope.

Outcome

Assumptions become geographies

Three defensible routing rules over identical data produced materially different maps of where the city’s waste burden falls.

Some structure is fixed

Corridors that persist across all methods mark where facility siting and network topology constrain the outcome regardless of routing logic.

Opacity has method consequences

Where operational systems are not public, modeling choices quietly become analytical claims. Naming them is part of the work.

The approach generalizes to any partially inaccessible infrastructure: build several defensible models, compare them honestly, and treat the divergence between them as evidence about the method. The geographies that become visible are inseparable from the assumptions used to construct them.