Assessing Glacier Extent Changes through Machine Learning Algorithms and Remote Sensing Data
Vanina Fissore, Lorenza Ranaldi, Davide Lisi, Piero Boccardo, Alessandro La Rocca, Mirko Frigerio, Daniele Sanmartino
Glaciers are critical elements in the Earth’s climate system, and can be considered as sensitive indicators of climate change. Glaciers store significant amounts of freshwater, which is essential for animal and human consumption and activities like industry and agriculture. Furthermore, glaciers have a significant impact on the hydrological cycle, and their melting also contributes to rising sea levels. Understanding and monitoring glacier extent changes is critical to informing climate policies, assessing natural hazards and safeguarding global water resources. Nowadays, remote sensing technology is a proved and widely adopted source of information in this sense.
In this context, the proposed study aims to develop a regression model able to predict future changes in glacier extent, using supervised machine learning algorithms applied to open access medium and HR spatial resolution satellite data of the EU Copernicus programme. To achieve this objective, two machine learning models are developed. The first model is a segmentation model that employs a U-Net architecture, along with a final Conditional Random Field (CRF) module, to digitalize glaciers features from satellite images. The purpose of the segmentation model is to vastly expand the dataset required by the regression model, in terms of glacier surface values. In fact, this work presents an additional contribution in the form of a novel dataset consisting of time series of glaciers and snow extent. This dataset is generated using the best-performing segmentation model previously trained, applied to multiple glaciers, spanning a 30-year period and a consistent seasonal interval. To train the segmentation model, and to create the required ground truth images, the GLIMS initiative database is used again, while optical satellite images are obtained in part from Sentinel-2 data and in part from other publicly available datasets such as the "Hindu Kush Himalayas (HKH) glacier mapping dataset". The latter couples annotated glacier locations, which were produced by experts, with multispectral imagery from Landsat 7.
The second model is a multivariate regression model that seeks to identify the relationships between Land Surface Temperature (LST) and glacier/snow extent.
In order to train the models, two datasets are required. For the regression model and specifically LST, data from the Sentinel-3 SLSTR instrument, as well as data from the ESA Climate Change Initiative, which consolidates data from various satellites over the past 25 years, are utilized. Historical data on glacier extent and elevation is obtained from the "Glaciers elevation and mass change data from 1850 to present from the Fluctuations of Glaciers" database by the Copernicus Climate Change Service and datasets provided by the Global Land Ice Measurements from Space (GLIMS) initiative. Finally, both models are validated on testing data to assess their generalization capabilities and their performance on real-world cases. A subset of the segmentation dataset is kept aside to extrapolate metrics such as the Intersection-Over-Union (IoU), which allows to assess the accuracy of the results obtained and to make comparison with other architectures. For the regression model, error metrics such as the Root-Mean Squared Error (RMSE) are considered to assess the model performance. The results of the study are expected to provide insights that will enhance the monitoring efforts of glacial features and provide useful information about the impact of climate change on glaciers worldwide.