Title: Reevaluating Amortized Neural Representations for High-Resolution Terrain Elevation Data

Abstract

Implicit neural representations (INRs) characterize signals as continuous functions mapping coordinates to values. In the context of terrain elevation, this approach facilitates the computation of analytic derivatives, enables decoding at arbitrary resolutions, and provides a smooth model of the underlying heightfield. Nevertheless, the strategy of training and maintaining a distinct INR for each individual tile fails to scale effectively when applied to expansive terrain datasets. Amortized neural representations address this scalability issue by employing a shared network architecture: a novel tile is encoded into a compact per-tile payload, which is then reconstructed by a shared decoder to generate the heightfield. While many existing techniques utilize hypernetworks to predict the payload in a single forward pass, other approaches determine the payload via brief, per-tile optimization processes. Initially designed for natural imagery, the applicability of these methods to terrain heightfields has not been fully established.

To address this uncertainty, we present a controlled benchmark utilizing a 1 m/pixel terrain dataset, where we assess three representative methods under a consistent protocol. Our analysis reveals a distinct cross-domain performance gap. In response, we introduce HUVR+SIREN, a hypernetwork that enhances the top-performing benchmark method, HUVR, by substituting its coordinate decoder with one that is smooth and analytically differentiable. This modification achieves superior fidelity for both height and derivatives on the benchmark, while requiring no extra per-tile storage and incurring lower decoding costs. Furthermore, HUVR+SIREN supports aggressive post-training quantization with minimal impact on quality, resulting in a highly compact neural format for terrain data. Additional ablation studies and diagnostics clarify which design elements are transferable to the terrain domain, indicating that the per-tile bottleneck is already approaching its practical limit and that the remaining performance gap stems from the architectural design of the shared hypernetwork.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC