rnd-dem-imputation

Impute missing OSM building heights using ArcticDEM raster elevation models and zonal statistics.

Overview

Many OpenStreetMap buildings lack height or level tags. This module fills those gaps by sampling a Digital Elevation Model (DEM) — primarily ArcticDEM — and deriving absolute building heights via terrain correction. It supports multiple imputation and terrain correction strategies, automatic tile fetching from NASA Earthdata, and optional orthometric height conversion.

raw DEM sample − terrain elevation = absolute building height

Features

• Three imputation strategies — max (default), direct, and coefficient (deprecated)
• Multiple DEM sampling modes — mean, max, median, percentile (decile), and more
• Terrain correction — buffer ring, elevation API (Google), or NASA DTM (SRTM / NASADEM)
• Automatic ArcticDEM tile fetching via pystac-client + NASA Earthdata auth
• Orthometric conversion — EGM2008 via pyproj + proj-data (enabled by default)
• Spatial interpolation of correction coefficients — IDW, KNN, and others (for coefficient strategy)
• Visualization tools — 3-D building extrusion and rendering for manual QA

Project Structure

rnd-dem-imputation/
├── src/height_imputation/                  # Core library
│   ├── height_imputer.py     # HeightImputer — main entry point
│   ├── arctic_dem.py         # ArcticDEM tile fetch & orthometric reproject
│   ├── dem_sampler.py        # Zonal statistics over building polygons
│   └── terrain_correction.py # TerrainCorrector strategies
├── visualization/        # 3-D extrusion & rendering helpers (manual QA)
├── tests/                # pytest test suite
├── environment.yml       # Conda/micromamba environment spec
└── setup.sh              # One-command environment setup

Quick Start

# 1. Create the environment (micromamba)
bash setup.sh

# 2. Activate
micromamba activate rnd-dem-imputation

Basic usage

import geopandas as gpd
from src.height_imputer import HeightImputer

buildings = gpd.read_file("buildings.geojson")

imputer = HeightImputer(
    dem_sampling_strategy="max",
    terrain_correction_strategy="buffer",
    imputation_strategy="max",          # recommended default
    arctic_dem_resolution="2m",
    earthdata_token="YOUR_TOKEN",
    cache_dir="cache/",
    verbose=True,
)

buildings_with_heights = imputer.impute(buildings)

Imputation Strategies

Strategy	Description
max (default)	height = nanmax(osm_height, levels × floor_height, dem_height) — takes the best available estimate
direct	Uses abs_dem_height as-is, no coefficient scaling
coefficient (deprecated)	Computes k = osm_height / abs_dem_height for tagged buildings, spatially interpolates k, then estimates heights for untagged buildings

Absolute DEM building height is always derived as:

abs_dem_height = raw_dem_sample − terrain_elevation

When use_orthometric=True (default), the DEM is converted to EGM2008 orthometric heights before sampling.

Terrain Correction Strategies

Strategy	Description
buffer	Aggregates DEM values in a ring buffer around each building footprint
elevation_api	Queries Google Elevation API for ground elevation
nasa_dtm	Uses SRTM or NASADEM as a reference terrain surface

Key Parameters

Parameter	Default	Description
dem_sampling_strategy	"max"	Zonal statistic for DEM sampling
dem_decile	90	Percentile for "decile" sampling
terrain_correction_strategy	"buffer"	Terrain correction method
buffer_radius	10	Ring-buffer width in metres
imputation_strategy	"max"	Height imputation strategy
floor_height	3.0	Metres per floor (for building:levels fallback)
use_orthometric	True	Convert DEM to EGM2008 before sampling
arctic_dem_resolution	"2m"	ArcticDEM mosaic resolution
cache_dir	None	Cache directory for DEM tiles

Requirements

• Python 3.11
• GDAL ≥ 3.7
• proj-data with EGM2008 grid (required for orthometric conversion)
• NASA Earthdata account + token (for ArcticDEM tile download)
• Google Elevation API key (optional, for elevation_api terrain correction)

See environment.yml for the full dependency list.

Running Tests

pytest tests/

License

Internal research module — see repository license for terms.