Corrected Antarctic Ice Sheet Surface Melt Estimates from Deep Learning, 2001-2016

A deep learning model and corrected surface melt time series for the Larsen Ice Shelf, Antarctica. The data includes a trained Multi-Layer Perceptron model and daily corrected melt estimates for three automatic weather station locations during austral summers from 2001 to 2016. The model was developed by researchers at Utrecht University as a proof-of-concept to improve regional climate model outputs using satellite and station data.

Use Cases

• Training deep learning models for climate variable correction based on albedo, temperature, and radiation inputs.
• Validating regional climate model (RACMO2) outputs against a machine learning-corrected benchmark.
• Analyzing temporal trends in Antarctic surface melt using a 16-year daily time series.
• Studying the relationship between surface albedo, radiation, and meltwater production on ice shelves.

Strengths

• Includes a pre-trained deep learning model (MLP) ready for application.
• Provides a 16-year daily time series (2001-2016) for three distinct locations.
• Corrected data is based on nine specific input variables including albedo, temperature, and radiation.

Limitations

• Data coverage is limited to three specific points on the Larsen Ice Shelf, not the entire continent.
• Corrected data are model outputs, not direct in-situ observations.
• Column-level documentation for the Excel files is absent; field semantics must be inferred from the description.

Provenance

Source

Utrecht University, associated with a 2021 preprint in The Cryosphere.

Collection Method

Correction of RACMO2 regional climate model simulations using a deep learning model trained on satellite and station data.

Time Range

2001 to 2016 (austral summers)

Geography

Larsen Ice Shelf, Antarctica (specific pixels corresponding to AWS 14, 17, and 18)