Periodic Reporting for period 1 - SKYNET (Estimating the ice volume of Earth's glaciers via Artificial Intelligence and remote sensing)
Periodo di rendicontazione: 2023-10-01 al 2025-09-30
Glaciers are shrinking at an accelerating pace as the planet warms. This melting contributes to sea-level rise and threatens freshwater resources for almost two billion people who depend on glacier-fed rivers around the world. The consequences range from coastal flooding to changes in water availability for agriculture, energy, and human consumption. Understanding how much ice remains in the world’s glaciers is therefore critical for anticipating and adapting to the impacts of climate change.
Yet, despite decades of research and direct measurements of glacier ice thickness, our knowledge of global glacier ice volume is still highly uncertain. Out of more than 216,000 glaciers on Earth, only about 3,000 have been directly surveyed with ice thickness measurements. Measuring ice thickness requires expensive and logistically challenging surveys, which are often limited to a few easily accessible regions. As a result, global estimates rely on models. While many different models have been developed and have advanced our knowledge of glacier ice volumes, their uncertainty is still often high, and in many cases, errors can exceed 30 to 40 percent, leading to large discrepancies in total ice volume estimates among different models.
At the same time, satellite technology has revolutionized the way we observe glaciers. High-resolution measurements of ice velocity, surface elevation, and mass change are now available globally from satellite missions. These vast datasets open new opportunities to infer ice thickness using modern data-driven techniques. Machine learning, in particular, offers a powerful way to identify patterns and relationships within complex datasets that traditional models cannot easily capture.
The SKYNET project builds on this potential. Its goal is to create a global machine learning framework capable of leveraging the wealth of observational data available to us from decades of surveys, by predicting ice thickness at any glacier on Earth, including in regions where no direct measurements exist.
The project pursues three main objectives:
1. To develop a global machine learning model to estimate glacier ice thickness, for every existing glacier on Earth.
2. To validate the model using all available observational measurements, compare it with existing approaches, and evaluate the strengths and weaknesses of the machine learning framework.
3. To generate global thickness maps for every glacier on Earth, openly available for researchers, policy-makers, and climate modelers alike.
By achieving these goals, SKYNET will help to refine our knowledge of global glacier ice volume. The project contributes directly to the objectives of the Paris Agreement, the EU Green Deal, and the IPCC’s priorities on cryosphere monitoring. More accurate knowledge of glacier ice volumes will support better projections of sea-level rise, improve regional water management planning, and strengthen our collective capacity to respond to climate change.
Ultimately, the SKYNET project aims to introduce advanced machine learning techniques into glacier modeling, providing additional tools and knowledge needed to understand the changing cryosphere in the 21st century.
Yet, despite decades of research and direct measurements of glacier ice thickness, our knowledge of global glacier ice volume is still highly uncertain. Out of more than 216,000 glaciers on Earth, only about 3,000 have been directly surveyed with ice thickness measurements. Measuring ice thickness requires expensive and logistically challenging surveys, which are often limited to a few easily accessible regions. As a result, global estimates rely on models. While many different models have been developed and have advanced our knowledge of glacier ice volumes, their uncertainty is still often high, and in many cases, errors can exceed 30 to 40 percent, leading to large discrepancies in total ice volume estimates among different models.
At the same time, satellite technology has revolutionized the way we observe glaciers. High-resolution measurements of ice velocity, surface elevation, and mass change are now available globally from satellite missions. These vast datasets open new opportunities to infer ice thickness using modern data-driven techniques. Machine learning, in particular, offers a powerful way to identify patterns and relationships within complex datasets that traditional models cannot easily capture.
The SKYNET project builds on this potential. Its goal is to create a global machine learning framework capable of leveraging the wealth of observational data available to us from decades of surveys, by predicting ice thickness at any glacier on Earth, including in regions where no direct measurements exist.
The project pursues three main objectives:
1. To develop a global machine learning model to estimate glacier ice thickness, for every existing glacier on Earth.
2. To validate the model using all available observational measurements, compare it with existing approaches, and evaluate the strengths and weaknesses of the machine learning framework.
3. To generate global thickness maps for every glacier on Earth, openly available for researchers, policy-makers, and climate modelers alike.
By achieving these goals, SKYNET will help to refine our knowledge of global glacier ice volume. The project contributes directly to the objectives of the Paris Agreement, the EU Green Deal, and the IPCC’s priorities on cryosphere monitoring. More accurate knowledge of glacier ice volumes will support better projections of sea-level rise, improve regional water management planning, and strengthen our collective capacity to respond to climate change.
Ultimately, the SKYNET project aims to introduce advanced machine learning techniques into glacier modeling, providing additional tools and knowledge needed to understand the changing cryosphere in the 21st century.
Following the start of the project, a secondment at the University of California, Irvine (UCI) was undertaken from 2023 to 2025. The main focus was to identify the most suitable machine learning architecture for a glacier thickness inversion problem. After evaluating several approaches, the choice converged on a fully supervised gradient-boosted decision tree algorithm, which are best handled using tabular learning methods rather than image-based architectures. This decision was motivated by the availability of millions of publicly accessible ice thickness measurements from ground penetrating radar surveys around the world, which are tabular in nature.
The second phase of the secondment focused on defining the optimal set of input features for the model training. These included the TanDEM-X Edited DEM v1, surface ice velocity, surface mass balance, temperature fields, and other geometrical attributes. A dedicated data pipeline was developed to generate a comprehensive training dataset, comprising millions of feature rows paired with corresponding thickness measurements.
The model was trained using a Bayesian optimization framework for hyperparameter tuning. Its performance was evaluated and validated on selected glaciers to assess generalization beyond the training domain. The resulting model, named IceBoost, integrates two advanced gradient boosting frameworks - XGBoost and CatBoost - and is trained on 3.7 million observed thickness measurements using 39 geophysical and climate-related inputs.
The model has demonstrated robust performance when compared to measurements, achieving up to 40% lower errors in polar regions compared to state-of-the-art modeling approaches. These results pave the way for the full-scale global deployment of the model and generation of glacier thickness maps for all existing glaciers, which is currently underway.
The second phase of the secondment focused on defining the optimal set of input features for the model training. These included the TanDEM-X Edited DEM v1, surface ice velocity, surface mass balance, temperature fields, and other geometrical attributes. A dedicated data pipeline was developed to generate a comprehensive training dataset, comprising millions of feature rows paired with corresponding thickness measurements.
The model was trained using a Bayesian optimization framework for hyperparameter tuning. Its performance was evaluated and validated on selected glaciers to assess generalization beyond the training domain. The resulting model, named IceBoost, integrates two advanced gradient boosting frameworks - XGBoost and CatBoost - and is trained on 3.7 million observed thickness measurements using 39 geophysical and climate-related inputs.
The model has demonstrated robust performance when compared to measurements, achieving up to 40% lower errors in polar regions compared to state-of-the-art modeling approaches. These results pave the way for the full-scale global deployment of the model and generation of glacier thickness maps for all existing glaciers, which is currently underway.
The project will deliver updated estimates of the volumes of all glaciers worldwide, providing essential information to refine projections of glacier contributions to sea-level rise by 2100 and to improve freshwater resource assessments in critical regions such as South America and the Himalayas. In addition, it will serve as a proof of concept demonstrating the potential of machine learning to tackle complex glaciological problems, offering innovative tools to deepen our understanding of a rapidly changing, warming world. Importantly, the project will also emphasize the value of openly accessible datasets and reproducible algorithms.