A Comparative Study of Machine Learning and Deep Learning Models for Long-Term Snow Depth Inversion.

Area of Science:

• Hydrology
• Climatology
• Environmental Science

Background:

• Snow depth is vital for understanding snow dynamics, water resources, and climate.
• Accurate snow depth data is essential for hydrological modeling, climate research, and disaster management in cold regions.

Purpose of the Study:

• To compare the performance of machine learning and deep learning models for daily snow depth retrieval.
• To evaluate the impact of meteorological factors, lagged snow depth, and physical snow parameters on prediction accuracy.

Main Methods:

• Utilized long-term daily meteorological data (1961-2015) from two stations in China.
• Integrated ERA5-Land reanalysis data for snow density and albedo.
• Compared three machine learning (XGBoost, Random Forest, SVM) and three deep learning (1D CNN, LSTM, 1D CNN-LSTM) models.
• Designed four feature combination schemes to assess variable importance.

Main Results:

• The first-order lagged snow depth was the most critical predictor for both model types.
• Machine learning models outperformed deep learning models, with XGBoost achieving the highest accuracy (R²=0.989, RMSE=1.19 cm).
• 1D CNN showed strong performance among deep learning models (R²=0.9878, RMSE=1.26 cm), but LSTM and hybrid models did not offer significant advantages.
• ERA5-Land snow physical parameters did not substantially improve prediction accuracy.

Conclusions:

• A prediction framework using historical snow depth and meteorological data is robust and effective.
• Model complexity does not directly correlate with predictive performance.
• Findings provide a robust solution for long-term snow depth reconstruction and support cryospheric process simulation and climate change studies.