Mapping vegetation with remote sensing and GIS data using object-based analysis and machine learning algorithms

Abstract

Remote sensing technology is an efficient tool for various practical applications of environmental resources management. Advances in this technology include the diverse range of high quality data sources and image analysis techniques. Object-based image analysis (OBIA) and machine learning algorithms are recent advances, which this thesis evaluates.

OBIA and machine learning algorithms are first tested using a combination of multiple datasets for identifying individual tree species. These datasets include Quickbird, LiDAR, and GIS derived terrain data. Improvements in tree species classification were obtained and the best data combination was terrain context (based on slope, elevation, and wetness), tree height, canopy shape, and branch density (based on LiDAR return intensity).

The availability of a range of classifiers and different data pre-processing techniques adds to the complexity of image analysis. The combinations of these techniques result in a large number of potential outcomes and these need to be evaluated. Therefore, the second part of this research investigated and compared tree species classification performance for different methods (Naïve Bayes - NB , Logistic Regression - LR, Random Forest - RF, and Support Vector Machine - SVM), combined with various dimensionality reduction (DR) methods (Correlation-based feature selection filter, Information Gain, Wrapper methods, and Principal Component Analysis). When DR was used prior to classification, only the NB classifier had a significant improvement in accuracy. SVM and RF had the best classification accuracy, and this was achieved without DR.

The final part of this thesis demonstrates a new method using OBIA for mapping the biomass change of mangrove forests in Vietnam between 2000 and 2011 from SPOT images. First, three different mangrove associations were identified using two levels of image segmentation followed by a SVM classifier and a range of spectral, texture and GIS information for classification. The RF regression model that integrated spectral, vegetation association type, texture, and vegetation indices obtained the highest accuracy.