Snowlaps: A Deep Learning Emulator for Biosnicar

snowlaps is a Python package that provides a fast, flexible alternative to the full biosnicar radiative transfer model. Built on a deep learning emulator trained on biosnicar v2.1 outputs, snowlaps enables rapid prediction and inversion of snow spectral albedo, making it a powerful tool for researchers and practitioners working with snow surface properties and light absorbing particles (LAPs). It is a lightweight and fast alternative to the full model for users that only need the core features of biosnicar.

What is Snowlaps?

snowlaps was originally developed to study the impact of different LAPs on snow spectral albedo in Southern Norway. It has since evolved into a general-purpose library with two main use cases:

• Forward mode: Predict snow spectral albedo from prescribed surface properties, serving as a fast alternative to running the full biosnicar model.
• Inverse mode: Infer surface properties from observed snow spectral albedo, enabling efficient parameter estimation and data analysis.

How Does Snowlaps Relate to Biosnicar?

The core of snowlaps is a deep learning emulator trained on a large set of biosnicar v2.1 simulations. This means that snowlaps can reproduce biosnicar’s outputs with high fidelity, but at a fraction of the computational cost. snowlaps is ideal for applications that require many model evaluations, such as sensitivity analyses, uncertainty quantification, or real-time data processing. It is likely to be more attractive than the ful model for many users, but power users that need to use the advanced features and/or full parameter space offered by biosnicar or want to build features on top of the core biosnicar functions should use the full model.

Why Use Snowlaps?

• Speed: The deep learning emulator provides near-instantaneous predictions, compared to the much longer runtimes of the full biosnicar model.
• Convenience: snowlaps can be used as a Python package or via an interactive Streamlit app, making it accessible for both scripting and exploratory analysis.
• Versatility: Supports both forward and inverse modeling, allowing users to predict albedo from surface properties or infer properties from measurements.
• Reproducibility: Trained directly on biosnicar outputs, ensuring consistency with the original model.

Usage

Installation

You can install snowlaps using conda and pip:

1. clone repository in the folder of your choice

git clone git@github.com:openosmia/snowlaps.git

2. move into snowlaps directory

cd snowlaps

3. create conda environment

conda env create -f environment.yml

4. activate conda environment

conda activate snowlaps

5. install snowlaps

pip install -e .

Run

snowlaps can be used in two main ways:

• As a Python package: Integrate snowlaps into your own scripts and workflows.
• Via the Streamlit app: Run an interactive web app for quick experimentation and visualization.

Example scripts are provided in the snowlaps/examples directory, including:

• Forward runs of the snowlaps emulator
• Inversion of hyperspectral albedo measurements
• Comparison of snowlaps and biosnicar predictions

Citation

If you use snowlaps in your research, please cite:

Chevrollier, L.-A., Wehrlé, A., Cook, J. M., Pirk, N., Benning, L. G., Anesio, A. M., and Tranter, M.: Separating the albedo reducing effect of different light absorbing particles on snow using deep learning, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2024-2583 (opens in a new tab), 2024.