Fast and Generalizable NeRF Architecture Selection for Satellite Scene Reconstruction

Abstract: Neural Radiance Fields (NeRF) have emerged as a powerful approach for photorealistic 3D reconstruction from multi-view images. However, deploying NeRF for satellite imagery remains challenging. Each scene requires individual training, and optimizing architectures via Neural Architecture Search (NAS) demands hours to days of GPU time. While existing approaches focus on architectural improvements, our SHAP analysis reveals that multi-view consistency, rather than model architecture, determines reconstruction quality. Based on this insight, we develop PreSCAN, a predictive framework that estimates NeRF quality prior to training using lightweight geometric and photometric descriptors. PreSCAN selects suitable architectures in < 30 seconds with < 1 dB prediction error, achieving 1000$\times$ speedup over NAS. We further demonstrate PreSCAN's deployment utility on edge platforms (Jetson Orin), where combining its predictions with offline cost profiling reduces inference power by 26% and latency by 43% with minimal quality loss. Experiments on DFC2019 datasets confirm that PreSCAN generalizes across diverse satellite scenes without retraining.

• The paper introduces PreSCAN, a linear predictor that uses multi-view consistency and scene descriptors to estimate NeRF PSNR before training, avoiding per-scene neural architecture search.
• The method achieves mostly sub-0.5 dB prediction error and completes selection in 28.3 seconds on average versus 9.7 hours for NAS, exceeding a 1000× speedup across held-out satellite scenes.
• The hardware-aware deployment study trades 0.79 dB of PSNR for 49% lower training power, 26% lower rendering power, 19% faster training, and 42% faster rendering on Jetson devices.

Motivation and problem statement

Neural Radiance Fields (NeRF) produce high-quality 3D reconstruction from multi-view imagery, but their application to satellite data is bottlenecked by two costs: per-scene training and per-scene architecture tuning. Neural Architecture Search (NAS) over NeRF configurations requires hours to days of GPU time per scene, which is prohibitive for mapping pipelines that must process thousands of locations. The paper challenges the assumption underlying most prior work — that architectural optimization is essential for reconstruction quality — and asks whether architecture selection can instead be performed before any training, using only cheap scene-level measurements.

The central finding: multi-view consistency dominates architecture

The authors conduct a SHAP-based feature importance analysis over 73 satellite scenes from the DFC2019 dataset (33 suburban Jacksonville scenes, 40 industrial Omaha scenes), using S-NeRF as the base model. The result is consistent across both regions individually and in combination: three scene-level descriptors — weighted inverse PSNR between view pairs ($w'$), mean photometric variance ($\overline{\text{Var}(I_i)}$), and view-direction cosine similarity ($\overline{\text{cos sim}}$) — dominate feature importance, while architectural parameters such as MLP depth, width, and ray-sample count have minimal influence.

This is the paper's strongest claim: multi-view geometric and photometric consistency, not model capacity, determines satellite NeRF reconstruction quality. Its implication is direct — if quality can be predicted from scene descriptors alone, per-scene NAS is unnecessary. The stability of feature rankings across JAX, OMA, and the merged dataset supports that this reflects a general property of satellite NeRF rather than dataset-specific artifacts, though the evidence base remains limited to two geographic regions and one primary model family.

The PreSCAN framework

PreSCAN concatenates four architectural parameters (number of layers $L$, hidden width $f$, samples per ray $N_{\text{sample}}$, sampling noise $\sigma_{\text{std}}$) with six scene descriptors into a linear regression predicting PSNR:

$$\text{PSNR}_{\text{pred}} = \boldsymbol{\beta}_{\text{arch}}^\top \mathbf{X}_{\text{arch}} + \boldsymbol{\beta}_{\text{scene}}^\top \mathbf{X}_{\text{scene}} + \beta_0$$

The scene descriptors are computed over valid image pairs defined by angular constraints on viewing angle and yaw differences, and include pair density $p(\phi)$, coverage score $C$, pairwise inverse-PSNR standard deviation, and view-angle statistics. The choice of an ordinary least squares estimator is deliberate and justified by three observations: SHAP analysis indicates approximately linear descriptor–quality relationships; linearity yields interpretable feature attributions; and linear models generalize under limited supervision without overfitting.

Training uses 3,650 scene–architecture pairs (top-50 architectures per scene selected via TPE-based NAS over a 144-configuration search space) from 64 scenes; OLS converges in under a minute. Prediction runs on a Tesla T4 — a weaker GPU than the A6000 used for NAS — in 28.3 seconds on average versus 9.7 hours for NAS, a speedup exceeding 1000×.

Predictive accuracy and generalization

On nine held-out test scenes spanning both regions, PreSCAN achieves MAE of 0.06–0.89 dB, with most predictions within 0.5 dB of ground truth. Worst cases (JAX-559 and OMA-331 at ≈0.89 dB) remain within the sub-dB target. Error distributions conditioned on architectural choices show additional structure: 128-feature models yield tighter error spreads than 64-feature variants, 10-layer networks show wider spread suggestive of overfitting, and increasing samples beyond 32 rays offers diminishing returns — indicating that model capacity matters more than sampling density, even if both are secondary to scene consistency.

Two robustness experiments strengthen the result:

• Sparse supervision: retraining on only 45 scenes preserves sub-dB error on most unseen scenes (best cases 0.01–0.08 dB). Notably, the largest error (OMA-244, 1.85 dB) occurs on a scene with extreme viewing geometry, and errors correlate with actual scene difficulty. This gives PreSCAN a second operational role: flagging problematic inputs for review.
• Model invariance: adding a model-variant term to the regression, the authors evaluate both S-NeRF and SatNeRF. MAE stays below 1 dB for both variants on all test scenes, supporting the claim that PreSCAN captures scene-driven difficulty rather than model-specific bias. This remains a modest validation, however — only two closely related satellite NeRF variants are tested.

Hardware-aware deployment case study

Because computational cost depends primarily on architecture rather than scene content, the authors decouple cost estimation from quality prediction. They profile all candidate architectures offline on Jetson Orin Nano and Orin AGX, recording training power, rendering power, training time, and rendering time in a lookup table. At deployment, PreSCAN predicts PSNR for all candidates, filters those below a quality threshold $\overline{\text{Var}(I_i)}$0, and selects the cheapest remaining option.

Against a best-accuracy baseline, this hardware-aware selection trades 0.79 dB of PSNR for substantial efficiency gains: 49% lower training power, 26% lower rendering power, 19% faster training, and 42% faster rendering. The pipeline requires no modification to PreSCAN itself and no hardware-in-the-loop testing during operation, making it plausible for onboard satellite processing where power budgets are strict.

Limitations and open questions

The paper's evidence carries several caveats stated or implicit in the evaluation. Generalization is demonstrated on a single dataset (DFC2019) covering two contrasting but geographically limited regions; behavior on other sensors, latitudes, or acquisition conditions is untested. Model invariance is shown for only S-NeRF and SatNeRF, leaving applicability to substantially different architectures (e.g., hash-encoding or grid-based methods such as SAT-NGP) open. The linear predictor assumes the approximately linear descriptor–quality relationship observed here holds more broadly, and the worst-case sparse-training error (1.85 dB) exceeds the sub-dB guarantee, suggesting extreme viewing geometries remain failure cases. Finally, the NAS ground truth itself derives from TPE search over a fixed 144-configuration space, so "suitable" architectures are relative to that space rather than globally optimal.

Conclusion

This paper reframes satellite NeRF architecture selection as a prediction problem rather than a search problem. Its principal contribution is empirical: SHAP analysis establishes that multi-view consistency features outweigh architectural parameters in determining reconstruction quality, and a deliberately simple linear predictor built on this insight delivers sub-dB PSNR prediction in under 30 seconds — a 1000× reduction versus NAS — while generalizing across regions, sparse supervision, and two NeRF variants. Combined with offline cost profiling, these predictions enable measurable power and latency savings on edge hardware with minimal quality loss. The framework's broader validity beyond DFC2019 scenes and beyond shadow-aware NeRF variants remains the key question left open by the evaluation.