Support vector regression snow-depth retrieval algorithm using passive microwave remote sensing data

doi:10.1016/j.rse.2018.03.008

2018-06-01

117

PRE

DOI：10.1016/j.rse.2018.03.008

原文链接

原文求助

摘要

En 中文

Snow cover is an informative indicator of climate change because it can affect local and regional surface energy and water balance, hydrological processes and climate, and ecosystem function. Passive microwave (PM) remote sensing data have long been used to retrieve snow depth and snow water equivalent with large uncertainties. The objective of this study is to develop snow-depth retrieval algorithm based on support vector regression (SVR) technique using PM remote sensing data and other auxiliary data. Ground-based daily snow depth data from 1223 stations across Eurasian continent were used to construct and validate the snow-depth retrieval algorithm. This SVR snow-depth retrieval algorithm partitioned three snow cover stages, and four land cover types then generated twelve snow-depth models for each phases. A non-linear regression method based on support vector regression (SVR) was used to retrieve snow depth with PM brightness temperatures, location (latitude and longitude), and terrain parameters (elevation) as input data and land cover as auxiliary data. In addition, we compared the performance of the SVR snow-depth retrieval algorithm with four alternative algorithms: the Chang algorithm, the Spectral Polarization Difference (SPD) algorithm, the Artificial/Neural Networks (ANN) and, an algorithm based on linear regression. Comparing results obtained from these five snow-depth retrieval algorithms against the ground-based daily snow depth data, the SVR snow-depth retrieval algorithm performs much superior with reduced uncertainties. We report the results aimed at evaluating the impact of the variation of snow cover stages and land cover types. The preliminary results suggest that the SVR snow-depth algorithm could detect deep snow with high accuracy and decrease the impact of saturation effects. These results suggest that the SVR snow-depth retrieval algorithm integrating PM remote sensing data and other auxiliary data (land cover types, location, terrain, snow cover stage with indirectly considering grain size variation) can be a more efficient and effective algorithm for retrieving snow depth and snow water equivalent over various scales.

Keyword:

Snow-depth retrieval

Snow cover stage

Land cover

Topography

Eurasia