Mapping Vegetation Cover Fractions using Brightness Corrected Hyperspectral Image Mosaics and Machine Learning Regression

Imaging spectroscopy together with machine learning regression is becoming increasingly viable for quantitative vegetation mapping and ecosystem assessment. However, large area assessments utilizing hyperspectral datasets require the use of image mosaics from multiple flightlines acquired over an extended period of time. This shifting of sun-sensor geometry can result in inconsistent across-track brightness gradients between the multiple flightlines of an image mosaic, thereby increasing spectral ambiguity between vegetation types and reducing the efficacy of mapping models. We mapped vegetation cover fractions in the San Francisco Bay Area using a hyperspectral image mosaic and Gaussian Process Regression (GPR) trained with synthetically mixed data from an image spectral library. The mosaic is composed of 11 EnMAP images simulated from the 2013 summer HyspIRI AVIRIS campaign. Due to severe across-track brightness gradients present in the imagery, various correction methods were explored. Results show that GPR is well suited for mapping vegetation fractions in heterogeneous landscapes. The fraction maps accurately depict transitions between different vegetation types across multiple ecosystems in central California. Empirical class-wise across track brightness correction produced more reliable fraction maps when compared to uncorrected imagery and other common brightness correction methods. Further, it also exhibited reduced confusion between vegetation types, particularly at large view angles. This work demonstrates the utility of GPR for mapping vegetation fractions and highlights the need to account for across track brightness gradients in regression modelling. This will be of particular importance when utilizing EnMAP's planned pointing capabilities for creating large area image mosaics.