Spatial probability membership value based ensemble: An improved tool for vegetation mapping

In the article a novel methodology is presented, aimed at performing supervised classification of hyperspectral data using spatially aware ensemble classification system. The work is a further development of the class-Probability Membership Value based Ensemble (PMVE) procedure, with the addition of spatial information into the originally spectral-based algorithm. The spatial context is incorporated by means of two approaches. The first one is a simple object recognition mechanism, and subsequent combination of the class probability values within each object. The second one is a local weighting approach, where this combination is achieved locally, considering a window around each pixel. Both spatial extensions to the original PMVE approach show moderate yet significant accuracy improvements and an increased robustness to local classification noise. The achieved overall and class-wise accuracy values show that these methodologies are useful to address generic vegetation mapping problems where discrimination of similar species is very difficult to achieve using more conventional hyperspectral data processing techniques.