Abstract
This paper introduces a spatial self-corrective learning (SCL) framework designed to transform sparse, "pencil-thin" satellite LiDAR measurements into high-resolution, full-coverage height maps. By integrating narrow-footprint ICESat-2 data with wide-coverage multispectral imagery, the SCL model employs confidence-based pseudo-interpolation and in-network regression to bridge massive data gaps. Experimental results demonstrate that SCL achieves significantly higher accuracy than traditional interpolation or standard deep learning methods.
Introduction
Monitoring Earth’s topography—especially in rapidly changing polar regions—is critical for understanding climate change. The ICESat-2 satellite provides incredibly precise surface height data, but it only "sees" what is directly beneath its narrow scan lines. This leaves massive "no-data" gaps, often kilometers wide, between observations.
While wide-coverage satellite imagery (like Landsat) can potentially fill these gaps, the relationship between an image's spectral pixels and actual ground height is complex. Spectral variations caused by soil type, moisture, or shadows often create noise that leads standard neural networks to produce "hallucinated" topography. Traditional spatial interpolation methods, such as Kriging, also struggle to capture the complex, nonlinear relationships between light and land.
Methodology
The proposed Spatial Self-Corrective Learning (SCL) framework treats height estimation as an iterative, self-refining process. Key innovations include:
Confidence-based Pseudo-Interpolation
The network does not just predict height; it generates a confidence score for every pixel. If the model is unsure about a prediction (e.g., due to unusual lighting), it "ignores" its own output and instead interpolates a value based on nearby high-confidence neighbors.
Recurrent Self-Refinement
The model passes its predictions back through the network multiple times, iteratively refining both the height map and the confidence scores to ensure spatial consistency.
Truth-Based Correction Layer ($F_R$)
A specialized ridge regression layer acts as a "sanity check." During the forward pass, it uses the limited ground-truth points available in the scan lines to solve a closed-form least-squares problem, instantly "warping" the rest of the tile's predictions to align with the known truth.
Base-Height Augmentation
To prevent the model from overfitting to specific absolute altitudes, the authors randomly shift the "base height" of training samples. This forces the network to focus on relative height variations—such as the shape of a mountain—rather than absolute elevation values.
Experimental Setup
The study utilizes several key remote sensing datasets:
- SRTM & IFSAR: Used as ground truth for training and evaluation across mid-latitude and Arctic regions.
- ICESat-2: The target dataset featuring narrow-footprint height observations.
- Landsat-7: Provides the 30m-resolution multispectral imagery used as the primary input feature.
Results and Discussion
The SCL framework was tested across diverse landscapes, including mountainous forests and snow-covered Arctic barrens.
Quantitative Results: SCL consistently outperformed baselines. On average, SCL achieved a symmetric Mean Absolute Percentage Error (sMAPE) of 0.123, compared to 0.164 for a standard U-Net and 0.193 for Kriging.
Qualitative Results: Visualizations show that SCL produces much smoother, more realistic terrains. While other models created "jagged" artifacts due to spectral noise, SCL successfully captured sharp mountain ridges and deep valleys while maintaining spatial consistency. Ablation studies confirmed that the recurrent refinement and the truth-based correction layer are critical to the architecture's performance.
Conclusion
The SCL framework provides a robust solution for generating full-coverage topography from sparse LiDAR data. By combining deep learning with spatial interpolation and on-the-fly correction, it overcomes the limitations of spectral noise and data sparsity. Future work will address challenges such as persistent cloud cover in polar regions and further evaluate the model's global robustness.
Summary updated