Super-resolved digital terrain models for GEDI geolocation correction: agreement with airborne LiDAR DTM-based corrections at the beam level, not the footprint level

Natsuki Yuda, Shinichiro ShirotaarXiv 2026 · arXiv

TL;DR

Researchers have developed a way to correct geolocation errors in GEDI spaceborne LiDAR data by using AI to "super-resolve" widely available 10m terrain models into 1m versions. While these AI-generated models are accurate enough to fix systematic, track-wide errors (beam-level), they still fall short for individual footprint corrections due to inherent errors in the original low-resolution data.

What Problem Does This Paper Address?

The Global Ecosystem Dynamics Investigation (GEDI) mission, operating from the International Space Station, provides critical 3D data on forest structure. However, it suffers from a significant hurdle: geolocation error. GEDI footprints (about 25m in diameter) can be off by roughly 10 meters horizontally. This misalignment makes it difficult to compare GEDI data with ground measurements or other satellite imagery, leading to errors in biomass and canopy height estimates.

Standard correction methods involve matching GEDI's ground elevation readings against high-resolution (1m) Digital Terrain Models (DTMs) derived from airborne LiDAR. The catch? These high-res DTMs are only available for a tiny fraction of the Earth's surface. This paper asks: Can we use deep learning to upscale widely available 10m DTMs to 1m and use them for geolocation correction instead?

Fig. 3.1: Schematic of GEDI observations. (a) The eight beams emitted from the ISS and the ground tracks they form. Thick lines are power beams, which originate from the high-output lasers, and thin lines are coverage beams. (b) A laser pulse illuminating one footprint (approximately 25 m in diameter) and the canopy. (c) The relationship between the received waveform of a footprint and the ground elevation (the elevation of the lowest mode), RH98, and RH100 derived from it. RH98 and RH100 are heights above the ground elevation.

Methods

The authors proposed a framework that integrates super-resolution (SR) with a two-stage geolocation correction process:

  1. Super-Resolution with "Mid-Fusion": They adapted a Modified Residual Channel Attention Network (RCAN) to upscale 10m DTMs to 1m. To reduce systematic bias, they introduced an auxiliary input: FABDEM, a global 30m DTM with forest and building heights removed. By fusing these two different data sources mid-network, the model learned to cancel out errors unique to each source.
  2. Two-Stage Correction:
    • Beam-level: A single horizontal shift is calculated for an entire ground track (beam).
  3. Footprint-level: Individual shifts are calculated for every single 25m footprint.
  4. Flow Accumulation: Instead of a simple "grid search" for the best fit, they used a hydrological "flow accumulation" algorithm. This treats the elevation difference map like a landscape and finds where "water" (probability) collects, making the correction more robust to local noise in the terrain data.

Datasets

The study focused on the steep, forested mountains of western Japan.

  • Primary Input: GSI 10m DTM (DEM10B), derived from topographic map contours.
  • Auxiliary Input: FABDEM (30m global DTM).
  • Training Target & Benchmark: GSI 1m DTM (DEM1A), derived from airborne LiDAR.
  • GEDI Data: Version 2 Level 2A products (ground elevation and RH98 canopy height).
  • Validation: 0.5m resolution Digital Canopy Height Models (DCHMs) from Hyogo and Kochi Prefectures.
Fig. 3.2: Arrangement of the patches used as Train, Valida- tion, and Test for the super-resolution model. The five test blocks for which DCHMs are available (5334_SE, 5234_NW, 5033_NE, 5033_SE, 4932_NE) are hatched. The hatched areas do not show the extents of the DCHMs themselves. The blank areas around the test blocks are buffers used for neither training nor validation.

Key Insights and Ideas

  • The Power of Error Cancellation: The "aha moment" is that super-resolved DTMs work remarkably well for beam-level corrections. Because a beam contains many footprints, the random errors in the AI-generated terrain tend to cancel each other out, leaving a reliable estimate of the systematic orbital shift.
  • The Footprint-Level Wall: Conversely, footprint-level correction fails to match the "gold standard" 1m DTM results. At the scale of a single 25m circle, the error in the super-resolved DTM is comparable to the elevation difference the correction is trying to resolve. The footprint simply "slides" into the nearest DTM error pit.
  • Garbage In, Garbage Out (Mostly): The authors found that most of the super-resolution error comes from the original 10m input, not the upscaling process itself. Improving geolocation correction globally will require better base-layer DTMs, not just fancier upscaling algorithms.
  • Flow Accumulation is Superior: The study confirms that flow accumulation is more stable than traditional grid searches, especially when the terrain lacks sharp, unique features.

Key Results

  • SR Accuracy: The trained model achieved an RMSE of 3.80m, a 33.9% improvement over standard bilinear interpolation.
  • Beam-Level Success: For 84.9% of the beams, the systematic shift calculated using the super-resolved DTM was within 5m of the shift calculated using the "true" 1m airborne LiDAR DTM.
  • Canopy Height Impact: After correction, the Mean Absolute Error (MAE) of canopy height was 5.24m with the generated DTM, compared to 5.02m with the 1m benchmark—a negligible difference of only 0.22m.
  • Bias Reduction: Fusing FABDEM reduced the vertical bias of the super-resolved model from 1.61m to just 0.10m.

Limitations and Future Work

The primary limitation is the availability of 10m DTMs. While more common than 1m data, many tropical and subtropical regions still lack free, high-quality 10m terrain models. Furthermore, the super-resolution model was trained specifically on Japanese terrain; its performance on different topographies (like the Amazon or the African savannah) remains unverified.

Future work suggested by the authors includes:

  • Developing "arbitrary-scale" super-resolution models that can handle any input resolution.
  • Incorporating optical satellite imagery (like Sentinel-2) as an additional auxiliary input to help the model "see" terrain features hidden in low-res DTMs.
  • Testing the transferability of these models across different global biomes.

Summary updated

Abstract

The Global Ecosystem Dynamics Investigation (GEDI) mission measures three-dimensional forest structure with a LiDAR aboard the International Space Station. GEDI observations have geolocation errors of about 10 m, which cause misalignment with other remote sensing data and reduce the accuracy of downstream tasks. Matching GEDI ground elevations against high-resolution digital terrain models (DTMs) from airborne LiDAR is one way to correct these errors, but such DTMs exist in few regions. To widen this method's applicability, we super-resolved a more widely available low-resolution DTM and used it for the matching. Our question is whether a super-resolved DTM yields the same correction as a 1 m-resolution airborne LiDAR DTM and, if not, why. In steep forests of western Japan, we super-resolved an existing 10 m-resolution DTM to 1 m. Improving its accuracy required fusion with a global 30 m-resolution DTM from a different source. Using the resulting DTM, we corrected GEDI observation positions first by a single shift per ground track, then per observation. For most tracks, the shifts nearly matched those from the airborne LiDAR DTM, likely because DTM errors canceled out across many observations. Per-observation corrections, however, did not match. At a single observation, the error does not cancel out and is too large for the correction. This error likely originates mainly from the input DTM. Therefore, even without a 1 m-resolution DTM, a roughly 10 m-resolution DTM, fused with a global DTM and super-resolved to 1 m, should yield track-wise corrections nearly matching those from a 1 m-resolution DTM.
BibTeX
@misc{yuda2026super,
  author = {Natsuki Yuda and Shinichiro Shirota},
  title = {{Super-resolved digital terrain models for GEDI geolocation correction: agreement with airborne LiDAR DTM-based corrections at the beam level, not the footprint level}},
  year = {2026},
  eprint = {2611.09854},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09854},
}
Figure