Prediction of Active Microwave Backscatter Over Snow-Covered Terrain Across Western Colorado Using a Land Surface...

Area of Science:

• Remote Sensing
• Snow Hydrology
• Machine Learning

Background:

• Accurate snowpack characterization is crucial for hydrological and climate studies.
• Satellite-based C-band backscatter offers potential for monitoring snow-covered terrain.
• Existing models often struggle to capture the complex electromagnetic response of snow.

Purpose of the Study:

• Develop a physically constrained support vector machine (SVM) for predicting C-band backscatter over snow.
• Utilize Sentinel-1 observations and Noah-MP land surface model outputs for training.
• Analyze the impact of training strategies on SVM prediction robustness.

Main Methods:

• Employed a physically constrained support vector machine (SVM) model.
• Used Sentinel-1 C-band backscatter as training targets.
• Incorporated geophysical variables from the Noah-MP land surface model as input.
• Investigated effects of training data (ascending/descending passes, time windows) and snow conditions (dry/wet).

Main Results:

• Combined ascending and descending overpasses increased prediction coverage to 15.2% but degraded accuracy.
• Longer training windows improved spatial coverage but introduced more random errors.
• Delineating dry versus wet snow pixels significantly improved predicted backscatter accuracy.

Conclusions:

• SVM prediction accuracy is strongly linked to the first-order physics of snow's electromagnetic response.
• Physically constrained SVM models show promise for snow remote sensing.
• Careful consideration of training data and snow conditions is essential for robust backscatter prediction.