Where the ground is hardest: machine-mapped impervious surface across the NJ Highlands, 2020
Impervious surface share across the New Jersey Highlands region (3,475 km2 of watershed lands supplying drinking water to much of northern New Jersey), aggregated to a hexagonal grid (1 km edge, roughly 2.6 km2 cells). The source layer was machine-mapped from NJGIN four-band aerial imagery with a Raster Vision semantic segmentation pipeline I built and operated for the NJ Highlands Council: 1,651 image tiles per imagery epoch, inference distributed across 10 to 20 AWS Spot instances, predictions vectorized and published by the Council as open data. Region-wide, 8.3 percent of Highlands land was impervious in 2020 (287 km2); the densest cell is 55 percent. Each hex carries land area, impervious area, and percent, so you can restyle or re-threshold it in any GIS.
Behind the work
Problem
Impervious cover is the most policy-relevant land indicator for watershed protection, but region-scale mapping of it by hand is impractical. This shows what an open-source ML pipeline can do with public imagery.
Role
Sole developer of the ML pipeline (model training, distributed inference, vector aggregation) as GIS consultant to the NJ Highlands Council. This hex aggregation is my own derivative work from the Council's published output layer.
Data sources
- NJGIN aerial imagery (2020 epoch)
- Highlands Impervious Surfaces 2020 and Highlands Boundary layers as published by the NJ Highlands Council on ArcGIS OpenData
Coordinate system
EPSG:4326 for delivery; analysis in EPSG:26918 (UTM 18N, meters)
Toolchain
Process notes
The published 2020 layer (815 dissolved multipolygons) was subdivided (ST_Subdivide) and intersected with a 1 km edge hex grid (ST_HexagonGrid) clipped to the Highlands boundary; percent = impervious area / land area within each cell. Cells under 1 ha of land (boundary slivers) dropped.
Validation
Percent bounded 0 to 54.9 across all 1,462 cells, none out of range; region-wide total (287.1 km2, 8.26 percent of land area) is consistent with published statewide land-use figures for the region. A 2007 vs 2020 change version was computed and deliberately NOT published: the only available 2007 source is raw model output (no curated 2007 layer exists on the portal), and raw-vs-curated comparison showed an implausible decrease, meaning the signal was curation difference, not ground change.
Data license
Published layers are NJ public data (copyright text: NJ Highlands Council). Output: CC BY 4.0