SatGeo-NeRF: Geometrically Regularized NeRF for Satellite Imagery

cs.CV Valentin Wagner, Sebastian Bullinger, Michael Arens, Rainer Stiefelhagen · Mar 23, 2026

What it does

Why it matters

The paper proposes three model-agnostic regularizers—gravity-aligned planarity, coarse-to-fine granularity masking, and depth supervision—to stabilize geometry learning. Experiments on the DFC2019 benchmark report 14% lower mean altitude...

Main concern

Community signal

0 up · 0 down

AI Review AI reviewed

Plain-language introduction

SatGeo-NeRF addresses wave-like geometric artifacts in satellite neural radiance fields caused by overfitting to multi-temporal imagery with varying lighting and transient objects. The paper proposes three model-agnostic regularizers—gravity-aligned planarity, coarse-to-fine granularity masking, and depth supervision—to stabilize geometry learning. Experiments on the DFC2019 benchmark report 14% lower mean altitude error relative to prior work, though this comparison relies on a reimplemented baseline that underperforms the original reported scores.

Critical review

Verdict

Bottom line

The paper presents a physically motivated solution to a genuine problem in satellite NeRFs: unconstrained geometry warping to compensate for photometric inconsistencies. The gravity-aligned planarity prior is well-suited to urban scenes, and the cross-ray gradient mechanism is technically sound. However, the central quantitative claim of 14% improvement over EO-NeRF rests on a self-acknowledged inferior reimplementation, complicating the assessment of true state-of-the-art status. The method is valuable for practitioners but requires stronger baseline validation.

“At the time of experiments, the official EO-NeRF implementation was not publicly available. We therefore rely on our reimplementation (EO-NeRF*). Our reproduced results do not reach the performance reported by Mari et al., (2023).”

Wagner et al. · Section 6.2

What holds up

The gravity-aligned planarity regularization is physically grounded and directly targets the observed pathology of wave-like artifacts on nominally flat urban surfaces. The coupling of adjacent rays via local surface approximation facilitates cross-ray gradient flow that standard per-pixel photometric losses cannot provide. The ablation study demonstrates that the three regularizers are complementary, with their combination achieving the lowest mean $MAE$ across all scenes.

“By explicitly computing surface normals with two auxiliary adjacent rays, the Gravity-Aligned Planarity loss $L_{planar}$ backpropagates through all samples along the three rays.”

Wagner et al. · Section 4.1

“The combination of all three regularization terms achieves the lowest mean MAE”

Wagner et al. · Table 2

Main concerns

The primary weakness is the reliance on EO-NeRF*—a reimplementation that the authors admit falls short of original reported performance—making the 14% improvement claim potentially inflated by implementation gaps rather than methodological gains. The evaluation is limited to four scenes from a single dataset (DFC2019), raising concerns about generalization to diverse urban morphologies or non-urban terrain where surfaces are not perpendicular to gravity. The $3.9\%$ PSNR drop on training views suggests a trade-off between geometric accuracy and photometric fidelity that could indicate over-regularization.

“This discrepancy in performance is in line with other reimplementations of EO-NeRF such as Behari et al., (2024).”

Wagner et al. · Section 6.2

“As the network is not able to overoptimize the geometry as freely, the PSNR value for the training views drops slightly by $3.9\%$.”

Wagner et al. · Section 6.2

Evidence and comparison

The evidence supports that geometric regularization reduces altitude error, with the ablation showing consistent improvements from each component. The comparison to EO-GS (Gaussian Splatting) is more convincing than to EO-NeRF since both are evaluated on the same reimplementation codebase. The claim of 11.4% improvement over EO-GS holds up, though both comparisons are limited to the DFC2019 Jacksonville scenes where flat urban structures dominate.

“Our proposed regularization are able to improve the MAE by a mean of $14.0\%$ on the DFC2019 dataset with only minimal impact on image render quality.”

Wagner et al. · Table 1

Reproducibility

The authors provide detailed hyperparameters ($300000$ iterations, batch size $1024$, $\lambda_{planar}=0.1$, $\lambda_{DS}=1000$) and specify the use of a reimplemented EO-NeRF backbone. However, there is no statement regarding code availability for SatGeo-NeRF itself. The reliance on bundle-adjusted RPC cameras and specific UTM coordinate alignment means reproduction requires substantial domain-specific preprocessing infrastructure. The computational optimization for adjacent ray sampling (using $50\%$ length) is well-documented and reduces the replication barrier.

“We employ the Adam optimizer with an initial learning rate of $5e-4$ and a batch size of $1024$.”

Wagner et al. · Section 5

“The Gravity-Aligned Planar Regularization $L_{planar}$ is turned on at epoch three once a coarse scene representation is learned, and its weight is empirically chosen as $\lambda_{planar}=0.1$.”

Wagner et al. · Section 5

Abstract

We present SatGeo-NeRF, a geometrically regularized NeRF for satellite imagery that mitigates overfitting-induced geometric artifacts observed in current state-of-the-art models using three model-agnostic regularizers. Gravity-Aligned Planarity Regularization aligns depth-inferred, approximated surface normals with the gravity axis to promote local planarity, coupling adjacent rays via a corresponding surface approximation to facilitate cross-ray gradient flow. Granularity Regularization enforces a coarse-to-fine geometry-learning scheme, and Depth-Supervised Regularization stabilizes early training for improved geometric accuracy. On the DFC2019 satellite reconstruction benchmark, SatGeo-NeRF improves the Mean Altitude Error by 13.9% and 11.7% relative to state-of-the-art baselines such as EO-NeRF and EO-GS.

Challenge the Review

Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.

Challenges are public to read, but only signed-in members can post them. Your challenge text is stored with your account for moderation, but usernames are not shown in the public thread.

No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.