University of Twente Student Theses


Machine learning regression for aboveground biomass estimation using PolSAR and PolInSAR data

Mukhopadyay, Ritwika (2020) Machine learning regression for aboveground biomass estimation using PolSAR and PolInSAR data.

Full text not available from this repository.

Full Text Status:Access to this publication is restricted
Abstract:Forests play a major role in the mitigation of climate change and in the regulation of the carbon cycle through carbon sequestration. The carbon is stored in leaves, branches, trunk and roots comprising of the above and below-ground biomass. Therefore, the biomass is a key biophysical parameter for the quality assessment of forests. The assessment of biomass helps in evaluating the carbon content and the variations in the carbon cycle. Techniques using GIS and remotely sensed data have immense contribution for monitoring worldwide health of forests. Microwave remote sensing has advantages in assessing biophysical parameters over other remote sensing techniques as it is capable of acquiring data in cloud cover and even at night time due to its long wavelengths. Microwave signal enables to retrieve scattering information from targets within each resolution cells. The present research work focuses on the machine learning (ML) regression-based retrieval and mapping of forest biophysical parameters of Malhan Forest Range of Dehradun Forest Division, using an image pair of fully polarimetric Radarsat-2 C-band data exploiting SAR Polarimetry (PolSAR) and Polarimetric SAR Interferometry (PolInSAR) techniques. Field inventory was executed in order to estimate field aboveground biomass (AGB) using the systematic sampling technique. The PolSAR backscatter components were retrieved using Yamaguchi four-component decomposition and H-alpha decomposition modelling approaches. The PolInSAR based coherence values were extracted for the different basis of polarization, namely, linear, Pauli and optimal. The retrieved information from the datasets was analysed with respect to the field estimated AGB and ground-to-volume scattering ratio, and then, the best combination of input variables for the regression algorithms was identified. The identified combination of input variables comprised of volume scattering information, sigma-nought HV backscatter coefficient, HV coherence and HV+VH coherence. The Random forest regression algorithm was trained and optimized using the field estimated AGB and the retrieved parameters in order to predict and map AGB for the entire study area. The feature importance was also evaluated for the input variables of the RFR algorithm. Through the analysis of the feature importance values, it was seen that the identified best combination of variables had higher sensitive in the prediction of AGB. A multiple-linear regression modelling was also used to develop a regression model for the estimation of AGB for the study area. A multi-collinearity test was implemented in order to reduce the interdependency and dimension of the independent variables. The validation and accuracy assessment was performed iteratively following which the results of the RFR algorithm showed a coefficient of determination (R2) value of 0.65 and a root mean squared error (RMSE) of 24.33 Mg ha-1, whereas, the MLR model results showed a coefficient of determination (R2) value of 0.54 and an RMSE of 33.05 Mg ha-1. Therefore, it can be inferred that the prediction of AGB through the RFR algorithm was found to be superior to that of MLR modelling approach, which can be used in combination with PolSAR and PolInSAR based components for the estimation of forest AGB. Finally, a comparative analysis was done for the two regression approaches based on the mapping of AGB for the entire forest.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page