Ensemble Learning From Synthetically Mixed Training Data for Quantifying Urban Land Cover With Support Vector Regression

Generating synthetically mixed data from library spectra provides a direct means to train empirical regression models for subpixel mapping. In order to best represent the subpixel composition of image data, the generation of synthetic mixtures must incorporate a multitude of mixing possibilities. This can lead to an excessive amount of training samples. We show that increasing mixing complexity in the training set improves model performance when quantifying urban land cover with support vector regression (SVR). To cope with the challenging increase in the number of training samples, we propose the use of ensemble learning based on bootstrap aggregation from synthetically mixed training data. The workflow is tested on simulated spaceborne imaging spectrometer data acquired over Berlin, Germany. Comparisons to SVR without bagging and multiple endmember spectral mixture analysis reveal the usefulness of the methodology for quantitative urban mapping.

I. Introduction

Exploiting imaging spectrometer data with machine learning has been demonstrated as an excellent choice for urban mapping [1]–[3]. On the one hand, the hyperspectral information enhances the separability of urban surface types due to material characteristic reflectance signatures in hundreds of contiguous spectral bands [4], [5]. On the other, machine learning algorithms effectively deal with high-dimensional image cubes and multimodal class distributions by fitting flexible, nonparametric, and nonlinear models without a priori assumptions on data distributions [6] . So far, most studies made use of airborne imaging spectrometer data with spatial resolutions below 5 m and hard per-pixel classifiers for urban land cover assessments [1] –[5]. With the forthcoming hyperspectral satellite missions Environmental Mapping and Analysis Program (EnMAP) [7] and Hyperspectral Infrared Imager (HyspIRI) [8], regional-scale urban mapping by means of 30 m resolution imaging spectrometer data will become possible. However, the dominance of spectrally mixed pixels calls for the use of quantitative techniques for subpixel mapping.