Fast and Generalizable NeRF Architecture Selection for Satellite Scene Reconstruction
Authors: Devjyoti Chakraborty, Zaki Sukma, Rakandhiya D. Rachmanto, Kriti Ghosh, In Kee Kim, Suchendra M. Bhandarkar, Lakshmish Ramaswamy, Nancy K. O'Hare, Deepak Mishra
Organizations: School of Computing, University of Georgia · Department of Geography, University of Georgia
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× 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.
Figures & tables
Figure 1: Overview of the PreSCAN framework. PreSCAN predicts suitable NeRF architectures before training by modeling the interaction between architectural parameters and scene descriptors. Using lightweight inputs (e.g., view alignment, sampling patterns), it enables fast ( < 30 sec.), generalizable, scalable architecture selection without per-scene retraining.
Figure 2: SHAP value analysis across JAX, OMA, and combined datasets. (a) ‘JAX scenes’: Inverse PSNR ( w′ ), photometric variance ( Var(Ii) ), and cosine similarity ( overlinecos sim ) are the most influential. (b) ‘OMA scenes’: The same top features are consistently identified, despite scene differences. (c) ‘Combined dataset’: Feature rankings remain stable, confirming their general importance across datasets.
Dataset
PSNR (Average)
PSNR (Max)
PSNR (Min)
JAX
19.63
23.03
14.16
OMA
17.18
23.06
12.21
Table 1: PSNR statistics across two datasets. The results show diversity in reconstruction difficulty and complexity.
Table 2: Comparison of NAS (on A6000 GPU) and PreSCAN (on Tesla T4, weaker GPU) prediction time. PreSCAN achieves > 1000 × speedup over NAS with comparable accuracy, enabling sub-minute ( < 30 sec.) prediction per scene.
Scene
Actual PSNR
Predict. PSNR
MAE ( Δ PSNR)
Std. Error
Bias
JAX-467
20.22
20.74
0.52
0.71
+0.52
JAX-474
20.15
20.57
0.42 (best)
0.31
+0.42
JAX-505
19.14
19.70
0.56
0.74
+0.56
JAX-559
19.06
19.95
0.89 (worst)
0.62
+0.89
OMA-258
19.10
19.04
0.07 (best)
0.13
-0.07
OMA-287
15.88
16.14
0.27
0.21
+0.27
Table 3: Comparison of actual vs. predicted average PSNR for scenes. We report MAE (absolute PSNR gap, Δ PSNR = ∣ Actual − Predicted ∣ ). Bold cells indicate low MAE (best predictions), and values marked as (worst) denote high error.
JAX (Unseen Scenes)
Scene
Act. PSNR
Pred. PSNR
MAE
Bias
JAX-276
19.88
19.24
0.64
-0.64
JAX-280
18.95
18.70
0.25
-0.25
JAX-314
18.53
18.60
0.08 (best)
+0.08
JAX-359
17.94
18.77
0.83 (worst)
+0.83
JAX-412
18.46
19.11
0.65
+0.65
Table 4: Scene-wise PSNR Generalization Comparison: Sparse Training (45 scenes) . Bold cells indicate low MAE (best), and values marked as (worst) denote high MAE error.
Figure 3: MAE ( Δ PSNR) distributions for different NeRF architectures across scenes. Each subplot shows prediction consistency per scene, with most errors < 1 dB dB and many clustering < 0.5 dB , demonstrating PreSCAN’s robust generalization.
Figure 4: Distribution of absolute prediction errors ( Δ PSNR) across various architectural parameters. (a) ‘Layers’: NeRF models with 6 and 8-layer yield tighter, more stable error distributions, while models with 10-layer shows higher spread and occasional outliers. (b) ‘Feature Dimensionality’: 128-feature models produce the most consistent predictions. (c) ‘Number of Samples’: Increasing the number of samples > 32 offers diminishing returns, with all settings clustering near 1 dB error.
Scene
S-NeRF MAE ( Δ PSNR)
SatNeRF MAE ( Δ PSNR)
JAX-467
0.79
0.82
JAX-474
0.55
0.77
JAX-505
0.61
0.91
JAX-559
0.03
0.57
OMA-258
0.19
0.02
OMA-287
0.37
0.56
Table 5: Comparison of MAE (Actual − Predicted PSNR) across NeRF variants. We report absolute PSNR error (Predicted - Actual) for S-NeRF and SatNeRF.
Figure 5: Comparison of power consumption and time between the most accurate model and the cost-efficient model on the edge device (Blue indicates baseline, red indicates PreSCAN chosen optimal). (a) Training power consumption, (b) Rendering power consumption, (c) Training time, (d) Rendering time, (e) Energy-PSNR tradeoff. PreSCAN with hardware-aware architecture selection achieves significant efficiency gains with minimal quality loss.
PSNR Diff
0.79 dB
MAE Diff
0.4 m
Power Consumption – Train
49%
Power Consumption – Render
26%
Time Reduction – Train
19%
Time Reduction – Render
42%
Table 6: Average percentage of benefit from hardware-aware architecture selection with the tradeoff in accuracy.