Scalable interpolation of satellite altimetry data with probabilistic machine learning

William K. Gregory, Ronald MacEachern, So Takao, Isobel R. Lawrence, Carmen Nab, Marc Peter Deisenroth, Michel Tsamados

TL;DR

This paper introduces GPSat, a new Python library that uses scalable Gaussian processes to efficiently interpolate satellite altimetry data. GPSat significantly speeds up the process of creating detailed maps of Arctic sea ice freeboard and sea-level anomalies, achieving a 504x speedup over previous methods while maintaining high accuracy. This advancement promises to unlock new insights into short-term ocean and sea ice variability.

What Problem Does This Paper Address?

Earth observation satellites, particularly altimeters, provide crucial data for understanding climate change, especially concerning sea-level rise and sea ice thickness. However, satellite altimetry data is inherently sparse due to the narrow tracks satellites follow. This sparsity limits our ability to study short-term (days to weeks) ocean and sea ice processes and hinders applications like initializing climate models or aiding maritime navigation. Existing interpolation methods, while useful, struggle to scale to the large datasets generated by modern satellites, often requiring significant computational resources or data sub-sampling, which can degrade accuracy. This creates a bottleneck in deriving high-resolution, timely data products.

Methods

The core of this work is the development of GPSat, an open-source Python library designed for efficient spatial and spatio-temporal interpolation of satellite altimetry data. GPSat leverages scalable Gaussian process (GP) regression techniques. Unlike traditional GP methods that compute a single model for the entire dataset, GPSat employs a local GP approach. This involves dividing the data into smaller, overlapping regions, each handled by a "local expert" GP model.

Key methodological innovations include:

  • Local Experts: Instead of a single global GP model, GPSat uses multiple local GP models, each trained on data within a defined spatial and temporal window around a prediction location.
  • Coarse Grid for Experts: To further improve scalability, GPSat defines these local experts on a coarser grid (200 km) than the final prediction grid. Predictions within an "inference domain" around each expert are made using that expert's model.
  • Weighted Averaging: Where inference domains overlap, predictions are combined using a weighted average based on distance to the expert locations.
  • GPU Acceleration and Modern ML Libraries: GPSat is built on GPflow and GPyTorch, enabling efficient linear algebra computations via GPUs and automatic differentiation, which significantly speeds up training and inference.
  • Inducing Point Methods (for future work): The paper also discusses the potential of sparse variational GP (SVGP) models, which use "inducing points" to approximate the data distribution, to further improve scalability for even finer resolutions.

Datasets

The study utilizes data from several satellite altimetry missions:

  • CryoSat-2 (CS2): Used for both radar freeboard and sea-level anomaly (SLA) data.
  • Sentinel-3A (S3A) and Sentinel-3B (S3B): Used for radar freeboard data.
  • NASA Operation IceBridge (OIB): Airborne data used for validating 5-km gridded radar freeboard predictions.

The paper also mentions the use of National Snow and Ice Data Center (NSIDC) NASA Team sea ice concentration data sets to define sea ice regions for interpolation.

Key Insights and Ideas

  • Scalability is Key: The primary insight is that by breaking down the interpolation problem into smaller, local GP models and leveraging modern ML tools (GPU, auto-diff), significant computational bottlenecks can be overcome.
  • Local Models Capture Non-Stationarity: The local GP approach is particularly well-suited for non-stationary data like sea ice and ocean fields, where statistical properties change across space and time.
  • Trade-off Between Resolution and Computation: While GPSat can achieve much finer resolutions (e.g., 5 km), there's a computational limit for a single GPU before memory becomes an issue, necessitating further research into methods like SVGP for ultra-high resolutions.
  • Improved Accuracy with Finer Grids: Higher resolution interpolation (5 km vs. 50 km) reveals finer spatial structures in sea ice freeboard, which are missed at coarser resolutions.
  • GPSat Outperforms Linear Regression: For both freeboard and SLA interpolation, GPSat demonstrates superior predictive skill compared to traditional linear regression, especially in capturing complex patterns and higher freeboard values.

Key Results

  • Computational Speedup: GPSat achieved a 504x computational speedup for 50 km-gridded Arctic sea ice radar freeboard interpolation compared to a previous method (G21N).
  • Accuracy: The derived freeboards from GPSat showed an average difference of less than 4 mm compared to the previous method.
  • High-Resolution Interpolation: GPSat successfully interpolated radar freeboards at 5 km resolution, showing strong agreement with airborne data (linear correlation of 0.66).
  • SLA Interpolation: Footprint-level SLA interpolation also demonstrated improvements in predictive skill over linear regression.
  • GPU vs. CPU: Moving from CPU to GPU implementations provided significant speedups (e.g., 4.8x for G21).
  • Local Expert Separation: The performance of GPSat is sensitive to the separation distance of local experts, with a 600 km separation yielding a 2270x speedup over G21N but slightly higher RMSD.

Limitations and Future Work

  • Computational Limits at Very Fine Resolutions: For along-track SLA interpolation, the memory limit of a single GPU restricts the size of the training domain. The authors suggest sparse variational GP (SVGP) models as a solution to scale to higher resolutions (e.g., 1 km for freeboard) and potentially improve SLA interpolation by incorporating temporal information.
  • Sensitivity to Expert Separation: The performance of GPSat can be sensitive to the chosen separation of local experts, suggesting this parameter needs careful testing for specific use cases.
  • Marginal Ice Zones: Predictions in marginal ice zones may carry higher uncertainty due to sparse sampling and dynamic ice conditions.
  • Noise Handling: At finer resolutions, GPSat can be prone to over-fitting noise. The authors applied a lower bound on the noise variance hyperparameter, but this needs rigorous testing for each use case.
  • Future Work: The authors plan to extend GPSat to the entire CryoSat-2 record, assess its impact on sea ice thickness trends, and explore SVGP models and parallelized GPU implementations for further speedups and improved accuracy. They also aim to conduct more extensive validation with airborne and ground-based data.

Summary updated

Abstract

We present GPSat; an open-source Python programming library for performing efficient interpolation of non-stationary satellite altimetry data, using scalable Gaussian process techniques. We use GPSat to generate complete maps of daily 50 km-gridded Arctic sea ice radar freeboard, and find that, relative to a previous interpolation scheme, GPSat offers a 504 × computational speedup, with less than 4 mm difference on the derived freeboards on average. We then demonstrate the scalability of GPSat through freeboard interpolation at 5 km resolution, and Sea-Level Anomalies (SLA) at the resolution of the altimeter footprint. Interpolated 5 km radar freeboards show strong agreement with airborne data (linear correlation of 0.66). Footprint-level SLA interpolation also shows improvements in predictive skill over linear regression. In this work, we suggest that GPSat could overcome the computational bottlenecks faced in many altimetry-based interpolation routines, and hence advance critical understanding of ocean and sea ice variability over short spatio-temporal scales.
BibTeX
@inproceedings{gregory2024scalable,
  author = {William K. Gregory and Ronald MacEachern and So Takao and Isobel R. Lawrence and Carmen Nab and Marc Peter Deisenroth and Michel Tsamados},
  title = {{Scalable interpolation of satellite altimetry data with probabilistic machine learning}},
  booktitle = {Nature Communications},
  year = {2024},
  doi = {10.1038/s41467-024-51900-x},
}
Figure