Optimizing Support Vector Machines with ISBA-A-gs Land Surface Variables as a Surrogate Model to Simulate ASCAT Derived Parameters

Optimizing Support Vector Machines with ISBA-A-gs Land Surface Variables as a Surrogate Model to Simulate ASCAT Derived Parameters

Kharagjitsing, Manish (TU Delft Civil Engineering and Geosciences)

Lhermitte, Stef (mentor)
Steele-Dunne, Susan (graduation committee)
Coenders-Gerrits, Miriam (graduation committee)

Delft University of Technology

Geoscience and Remote Sensing

2020-02-28

The TU-Wien developed a soil moisture retrieval algorithm that uses the incidence angle dependence of backscatter to obtain soil moisture estimates (Wagner et al., 1999). The core of this algorithm is a second order Taylor expansion with which the backscatter is normalized at a reference angle. Studies have shown that the first and second order derivative within this Taylor expansion, known as slope and curvature, are somehow related to the wet biomass and structure of vegetation. The general approach to forward model satellite observations with land surface variables in a data assimilation framework is through a radiative transfer model (Albergel et al., 2017). However, this requires plenty of assumptions about the vegetation canopy (such as stem height, shape, size, orientation etc.) and is therefore relatively inefficient for understanding the impact of soil moisture and vegetation dynamics on backscatter on a large scale. This study investigates the possibility of using support vector machines as a surrogate model instead of a radiative transfer model to link the TU-Wien normalized backscatter and slope to land surface variables soil moisture and leaf area index. The land surface variables are simulations from the CO2-responsive ISBA-A-gs land surface model. Support vector machines have the advantage of providing implicit kernel functions, which make them very useful for non-linear problems. The ISBA-A-gs data is provided by Météo-France. In total, 1324 support vector machines have been optimized through a cross validated grid search. The optimized hyperparameters were shown to have spatial consistency and look promising as an initial approach to forward modelling backscatter and slope. The SVM performances are further investigated through corresponding land cover types of grid points and the land surface variables.

ASCAT
Remote Sensing
Machine Learning
Support Vector Machine

http://resolver.tudelft.nl/uuid:2a9c3314-57b6-4471-9d81-b79199e5ffbd

2020-03-01

Student theses

master thesis

© 2020 Manish Kharagjitsing