JAX 004 postprocessed surface with regularized SAM building roofs and estimated ground

Satellite imagery. Urban geometry.

AgentDSM.

Building-aware
satellite reconstruction.

Semantic priors guide the reconstruction.
Scene-level evidence guides the decision.

View before & after

JAX_004 / JACKSONVILLE, FL

Static 3.0 / buildings + ground · 3× vertical scale

Buildings + estimated ground / display layer

Fitted roofsWallsEstimated ground

RECONSTRUCTION ACCURACY / BEFORE GEOMETRIC POSTPROCESSING

BUILDING-FOCUSED / JAX_004

0.844 → 0.806m

Building MAE · Static weight 2.0

GENERAL DSM / JAX_004

1.362 → 1.349m

Full-scene MAE · Staged schedule

SCENE-SPECIFIC DECISIONS

4evaluated AOIs

Recorded policies, regional metrics, audit trails

01 / RECONSTRUCTION

The surface. The difference.

One urban scene, two product objectives.
Recorded reconstruction results.

Building-focused DSM elevation using static semantic weight 2.0

Baseline EOGSStatic 2.050 m

Elevation (m)

Registered DSMs on the same grid; shared elevation color limits.

RECONSTRUCTION VARIANT

More weight on buildings.

Static semantic weight 2.0 gives the lowest building-region error among the evaluated JAX_004 candidates.

4.5%lower building MAE

Baseline versus selected policy, MAE in meters
MAE (m)	Baseline	Static 2.0

Full-scene MAE increases from 1.3625 m to 1.3757 m at this building-focused operating point.

DFC2019 · JAX_0045,000 training iterations0.5 m benchmark grid

View data provenance

02 / SEMANTIC PRIOR + POLICY CONTROL

From satellite pixels
to an inspectable product.

Building masks express what matters.
Validation determines which policy to keep.

Satellite view of JAX 004 with predicted SAM building masks in red — A real input view with predicted building masks. These image-space masks weight the photometric loss; they do not provide known heights.

Observe the scene

Multi-view satellite images, camera models, SAM mask coverage, and inventory metadata.

Reconstruct with a prior

Building-weighted Gaussian Splatting with baseline, static, and staged candidate schedules.

w(p) = 1 + (λ_b − 1) M_b(p)

Verify the objective

Compare full-scene, building, vegetation, and non-building MAE. Retain the validated policy for the requested product.

Deliver the evidence

Registered DSM, regional metrics, inventory, and decision records. Optional SAM-guided postprocessing produces regularized roofs and an estimated-ground display layer.

03 / RECORDED AGENT DECISIONS

A policy for each scene.

Semantic weighting helps JAX_004.
The other three AOIs retain their validated baseline.

Recorded policy selections. Scene buttons open their audit records.
Scene	SAM coverage	Building area	Components	Selected policy	Full MAE (m)

DECISION RECORD JAX_004

Staged reconstruction selected.

Audit JSON

Offline selection from completed candidate-validation metrics, not a learned policy evaluated on unseen scenes. Building area and component counts come from the benchmark CLS=6 layer (components ≥25 m²), not cadastral records or SAM predictions.

JAX_004 ablation: all evaluated operating points

JAX 004 ablation, MAE in meters
Policy	Full scene	Buildings	Vegetation	Non-building

04 / SAM-GUIDED POSTPROCESSING

Static 2.0 + SAM geometry.
Static 3.0.

A visual edition of Static 2.0.
SAM-guided roofs, sharper building edges.

Static 2.0 and Static 3.0 visual comparison with matching full-scene views and close-ups of SAM-guided building edges — Same grid and elevation colors. Static 3.0 refines SAM-supported roofs and building edges, preserving the original scene outside the footprints and a 1 m exterior edge band.

Before and after SAM-guided postprocessing: original DSM with vegetation on the left; regularized buildings and estimated ground on the right. Same camera and 3x vertical scale. — JAX_004: 11 projected SAM building-mask views support 59 retained components from Static 2.0. Geometric postprocessing regularizes their footprints and fits roof planes; the ground is estimated from the reconstructed DSM.

01 / FOOTPRINTS

SAM-guided boundaries

Multi-view mask fusion, contour simplification, and rectangular fitting where supported.

02 / BUILDINGS

Roof planes and walls

Robust planar fits retain building height structure and make boundaries explicit.

03 / CONTEXT

Estimated ground

Morphological terrain estimation replaces vegetation and other non-building relief.

Static 3.0 is a visualization-only postprocessing of Static 2.0, separate from the reconstruction results above. The buildings + ground view omits non-building relief, including vegetation and any buildings missed by the masks.

Static 3.0 display GeoTIFF Surface image Postprocessed GeoTIFF Processing record Compare processing variants

Reconstruction audit and original product sheet

Original JAX 004 reconstruction product sheet with audit records and benchmark-derived inventory — Original offline reconstruction report. This inventory uses the benchmark building layer and is separate from the 59 SAM-supported postprocessing components.

05 / THE RESEARCH

Agentic Building-Aware Satellite Gaussian Splatting for Auditable Urban DSM Reconstruction

Wentao Sun¹, Zhengsen Xu², Yiping Chen³, John S. Zelek¹, Jonathan Li¹

¹ University of Waterloo² University of Calgary³ Sun Yat-sen University

AgentDSM combines SAM-derived building priors with a reconstruction controller around Earth-observation Gaussian Splatting. It makes the building-versus-global accuracy trade-off explicit, selects scene-specific operating points using validation evidence, and exports DSM products with inspectable decision records.

Results CSV Data provenance

Citation

@unpublished{sun_agentdsm_2026,
  title = {Agentic Building-Aware Satellite Gaussian Splatting
           for Auditable Urban DSM Reconstruction},
  author = {Sun, Wentao and Xu, Zhengsen and Chen, Yiping
            and Zelek, John S. and Li, Jonathan},
  year = {2026},
  note = {Research manuscript},
  url = {https://w27sun.github.io/agentdsm/}
}