The article of Hans Roelofsen, Lammert Kooistra, Peter van Bodegom, Jochem Verrelst, Johan Krol and Jan-Philip Witte: Mapping a priori defined plant associations using remotely sensed vegetation characteristics has been published in Remote Sensing of Environment, Volume 140, January 2014, Pages 639–651.
AbstractIncorporation of a priori defined plant associations into remote sensing products is a major challenge that has only recently been confronted by the remote sensing community. We present an approach to map the spatial distribution of such associations by using plant indicator values (IVs) for salinity, moisture and nutrients as an intermediate between spectral reflectance and association occurrences. For a 12 km2 study site in the Netherlands, the relations between observed IVs at local vegetation plots and visible and near-infrared (VNIR) and short-wave infrared (SWIR) airborne reflectance data were modelled using Gaussian Process Regression (GPR) (R2 0.73, 0.64 and 0.76 for salinity, moisture and nutrients, respectively). These relations were applied to map IVs for the complete study site. Association occurrence probabilities were modelled as function of IVs using a large database of vegetation plots with known association and IVs. Using the mapped IVs, we calculated occurrence probabilities of 19 associations for each pixel, resulting in both a crisp association map with the most likely occurring association per pixel, as well as occurrence probability maps per association. Association occurrence predictions were assessed by a local vegetation expert, which revealed that the occurrences of associations situated at frequently predicted indicator value combinations were over predicted. This seems primarily due to biases in the GPR predicted IVs, resulting in associations with envelopes located in extreme ends of IVs being scarcely predicted.
Although the results of this particular study were not fully satisfactory, the method potentially offers several advantages compared to current vegetation classification techniques, like site-independent calibration of association probabilities, site-independent selection of associations and the provision of IV maps and occurrence probabilities per association. If the prediction of IVs can be improved, this method may thus provide a viable roadmap to bring a priori defined plant associations into the domain of remote sensing.
Keywords: Phytosociology; Coastal ecosystem; Gaussian Mixture Density Modelling; Gaussian Process Regression; Indicator values; Probability Density Function; Optical airborne remote sensing