Source-linked AI summary
Spatially explicit feature importance for building height estimation using research-access high-resolution SAR and optical sensors
Guilherme Iablonovski, Pierre-Louis Frison, Tatiana Silva da Silva
TL;DR
Individual-footprint building heights remain difficult to obtain in Global South cities because LiDAR and commercial very-high-resolution imagery are scarce or costly, while Sentinel-based products are too coarse. This study combines research-access TerraSAR-X, PlanetScope, and Sentinel-1 features in a geographically weighted random forest, achieving 5.34 m RMSE and R² = 0.756 against LiDAR while revealing context-dependent predictor dominance.
Problem
Building-height estimates remain limited by costly or scarce high-resolution data and by Sentinel products whose 10 m resolution cannot resolve individual building footprints.
Method
The study integrates TerraSAR-X, PlanetScope, and Sentinel-1-derived predictors in a geographically weighted random forest trained against LiDAR heights.
Results
5.34 m RMSE and R² = 0.756 were achieved, while predictor dominance varied systematically by building height and urban context.
Takeaways & Limitations
Research-access satellite products can support footprint-level height estimation and provide context-specific optioneering guidance beyond global pixel-level models.
Takeaways & Limitations
Whether the identified spatial predictor patterns generalize to other cities remains to be examined.
Abstract
from arXiv · showhide
Accurate building height information at the individual footprint scale is essential for material stock accounting and post-disaster damage assessments yet remains difficult to obtain at city scale in the Global South where airborne LiDAR coverage is rare and commercial very high-resolution imagery is cost-prohibitive or unavailable. While recent works have demonstrated building height estimation using freely available Sentinel imagery, the resolution ceiling of resulting products is still coarse for material stock analysis. This study incorporates products derived from data freely accessible under scientific research licenses, TerraSAR-X StripMap and PlanetScope, alongside Sentinel-1 to predict building heights in a large city in Brazil. To account for the spatial autocorrelation in the training set, features from all sources are integrated in a geographically weighted random forest model, returning an RMSE of 5.34 m and R2 of 0.756 against a LiDAR reference dataset. Local feature importance showed predictor dominance to vary consistently across intra-urban contexts, with footprint geometry dominating for low-rise buildings, shadow-derived height for taller and more isolated structures, and spectral reflectance for the tallest buildings in the set. Sentinel-1 backscatter and InSAR occupy complementary spatial niches, with no single sensor uniformly preferable across the set. Results provide optioneering guidance and insight over satellite-derived products predictive relevance in distinct contexts, which global machine learning or neural network models cannot offer.
I. INTRODUCTION
Building height information supports urban applications but is difficult to obtain where airborne LiDAR and very high-resolution satellite data are costly. This study addresses that gap by integrating research-access TerraSAR-X and PlanetScope imagery with open-access satellite information to estimate building heights in Porto Alegre, Brazil.
- Motivation: Building height data support applications including urban damage assessment after disasters.Remote-sensing approaches include airborne laser scanning, optical shadow analysis, interferometric SAR, and stereophotogrammetry.
- Data gap: Many existing building-height methods depend on costly airborne LiDAR or very high-resolution satellite imagery.These data constraints limit access to building-height information in many settings.
- Prior work: Open-access SAR and optical satellite pipelines have emerged to address the three-dimensional building-data gap.A Sentinel-1 and Sentinel-2 time-series support vector machine achieved an RMSE of 6.07 m at 10 m spatial resolution across Germany.
- Contribution: The proposed approach combines TerraSAR-X and PlanetScope imagery with open-access satellite-derived information for building-height prediction in Porto Alegre, Brazil.TerraSAR-X is available through the DLR science program, while PlanetScope is available through Planet's Education and Research Program.
II. METHODS · A. Reference Dataset creation
The reference dataset combines LiDAR-derived building heights with PlanetScope, TerraSAR-X, Sentinel-1, shadow, and footprint-geometry features extracted per building footprint. These data provide complementary optical, SAR, and geometric predictors at their native spatial scales.
- A. Reference Dataset creation: 2.2 points per m² minimum LiDAR density and 7.5 cm mean error support Porto Alegre’s building-height reference dataset.Reference height equals DSM minus DTM at each building footprint.
- A. Reference Dataset creation: PlanetScope 3 m, four-band surface-reflectance composites provide spectral indices and near-infrared GLCM texture features.Indices include NDVI, a proxy NDBI, and brightness; textures include contrast, entropy, energy, correlation, dissimilarity, and homogeneity.
- A. Reference Dataset creation: Shadow pixels darker than 0.85 across all bands yield shadow-length height estimates adjusted for solar incidence angle and azimuth.Estimates were capped at 3–110 m to remove implausible values caused by shadow occlusion.
- A. Reference Dataset creation: 177 m ambiguity-height TerraSAR-X StripMap interferometry produces a DSM through coregistration, filtering, unwrapping, and terrain correction.StripMap was selected for systematic acquisitions and wider swath width than higher-resolution modes.
- A. Reference Dataset creation: 20 m Sentinel-1 GRD imagery complements TerraSAR-X through neighborhood-level volumetric double-bounce signals from wall–ground interactions.All 2015 images covering the study area were processed to create the VVH indicator.
- A. Reference Dataset creation: Footprint attributes include area, perimeter, interbuilding distances, and neighboring-building counts within 25 m, 50 m, and 100 m.These variables characterize building context and remote-sensing signal quality.
- A. Reference Dataset creation: All feature sets undergo all-touched zonal-statistics extraction per footprint, producing a vector database of integrated variables.The extraction rule ensures adequate pixel coverage for small footprints.
B. Statistical Analysis
The study selected geographically weighted random forest regression after exploratory analysis identified spatial and distributional challenges. Its spatially local feature-importance framework supports context-specific assessment of predictor contributions while avoiding bias from unequal group sizes.
- Preliminary analysis: 66 explanatory variables across seven groups were assembled to predict LiDAR-derived building height, with preliminary analysis assessing spatial autocorrelation.The groups included optical spectral and texture, shadow-derived heights, InSAR-derived heights, backscatter intensity, and backscatter texture.
- Model selection: Geographically weighted regression was discarded because the dataset violated normality and multicollinearity assumptions, while geographically weighted random forest addressed spatial structure.Random forest avoids those assumptions but treats observations as spatially exchangeable, motivating the GWRF choice.
- Geographically weighted Random Forest: GWRF fits proximity-weighted local random forests for each building and combines local and global predictions.Bandwidth and the local-versus-global weighting were informed by incremental spatial autocorrelation analysis.
- Sampling: 33,212 buildings were sampled spatially to include every building at least 12m tall while preserving representative geographic coverage.Buildings taller than 12m comprised 4% of the stock and were disproportionately harder to predict than short buildings.
- Feature importance: Local feature importance was calculated separately for each building, then aggregated by group mean to reduce bias from groups containing more predictors.Random-forest importance measures split-based predictive contribution but does not reveal whether a predictor’s relationship with height is positive or negative.
C. Urban morphology features
The study computed seven urban morphology and contextual variables for each building to characterize which building types and urban contexts are best described by each predictor group.
- Urban morphology features: Seven variables captured terrain, footprint form, and neighborhood context for each building.The variables were altitude, slope, footprint orientation, elongation, vertex count, neighborhood built density, and height relative to the five nearest neighbors’ mean.
III. RESULTS
The geographically weighted random forest achieved strong footprint-scale height-estimation performance, while spatially explicit importance revealed systematic differences across building heights and urban contexts. These results support interpretable sensor optioneering using research-access remote sensing data.
- Model performance: 5.34 m RMSE and R² = 0.756 were achieved by the combined GWRF, outperforming the global-only and local-only components.The global-only component achieved RMSE = 5.76 m, R² = 0.716; the local-only component achieved RMSE = 5.53 m, R² = 0.739.
- Feature importance: Geometry dominated 32.4% of buildings, followed by shadow-derived height at 24.5% and PlanetScope spectral reflectance at 16.0%.PlanetScope texture, InSAR-derived height, Sentinel-1 texture, and Sentinel-1 backscatter dominated 11.9%, 9.0%, 3.5%, and 2.7%, respectively.
- Height-stratified importance: Geometry’s dominance declined from 42.9% of buildings under 6 m to 17.7% above 35 m, showing systematic variation in dominant predictors by building height.The passage identifies each building’s dominant group as the one with the highest mean importance.
- Methodological interpretation: InSAR-derived height dominated buildings at low altitude and slope, reflecting a methodological artifact from nDSM ground-surface mismatch in complex or rapidly varying terrain.The InSAR-derived nDSM differences the interferometric DSM against FABDEM.
- Height-stratified importance: PlanetScope spectral reflectance rose from 10% to 40% across height strata and became the largest group among the tallest buildings.Buildings dominated by spectral reflectance occurred in significantly denser environments than the other five groups.
- Morphological heterogeneity: All twelve morphology variables differed significantly across the seven dominant-group categories, with all p < 0.001.This result confirms the spatial heterogeneity hypothesis across multiple independent morphological dimensions.