What question did this study set out to answer?

The aim is to create a global daily product for snow water equivalent (SWE) using machine learning and in-situ measurements.

March 13, 2026Scientific Data0 citationsOpen Access

Global 0.25-degree gridded Snow water equivalent data derived from machine learning using in-situ measurements

JSJungho Seo MPMahdi Panahi JKJunsu Kim

Key Points

The aim is to create a global daily product for snow water equivalent (SWE) using machine learning and in-situ measurements.
Developed a machine learning-based SWE product at 0.25° resolution for 1980-2020.
Used k-means clustering to group in-situ SWE measurements into 14 clusters.
Trained a random forest model on 11,687 grid points using meteorological and terrain data.
Compared the model with ten reference datasets, including land-atmosphere reanalysis and remote sensing products.
Achieved an overall root mean square error (RMSE) of 10.33 mm and bias of -7.13 mm.
High accuracy in the Rocky Mountains, with RMSE of 7.30 mm and correlation coefficient of 0.98.
Showed agreement with Gamma airborne SWE over North America and similar spatial patterns to Andes Snow Reanalysis.

Abstract

In this study, we developed a machine learning-based global daily SWE product (SWEML) with 0.25° (~25 km) resolution for 1980-2020. Using k-means clustering, in-situ SWE measurements were grouped into 14 clusters, and a random forest model was trained on 11,687 grid points with meteorological forcing and terrain attributes. SWEML was compared with ten reference datasets representing diverse approaches, including land-atmosphere reanalysis without data assimilation (DA), systems incorporating DA, snow model simulations of varying complexity driven by reanalysis forcing with or without DA, and remote sensing products. The overall root mean square error (RMSE) and bias were 10.33 mm and -7.13 mm, respectively. Notably, SWEML achieved high accuracy in high-elevation regions such as the Rocky Mountains, with an RMSE of 7.30 mm and correlation coefficient of 0.98. It also agreed with the Gamma airborne SWE over North America and showed similar spatial patterns and peak SWE time series of the Andes Snow Reanalysis. These results highlight the robustness of SWEML in regions with and without training data.

AI에게 질문

Bookmark

View Full Paper