River floodplain vegetation classification using multi-temporal high-resolution colour infrared UAV imagery.

To evaluate floodplain functioning, monitoring of its vegetation is essential. Although airborne imagery is widely applied for this purpose, classification accuracy (CA) remains low for grassland (< 88%) and herbaceous vegetation (<57%) due to the spectral and structural similarity of these vegetation types. Increased availability of Unmanned Aerial Vehicles (UAV) allows low-cost production of high-resolution orthophotos and digital surface models (DSMs). Multi-temporal DSMs and orthophotos may be used as input for an improved classification methodology, using differences in phenological changes between vegetation types. The aim of this study was (1) to evaluate the improvement of the CA when using multi-temporal UAV-derived imagery and (2) to determine which layers of a multi-temporal imagery and derived DSMs yield an optimal balance between CA and acquisition effort. During six field surveys with six to ten weeks intervals over one year, a floodplain section along the lower Rhine, the Netherlands, was recorded with true-colour and false-colour imagery with a UAV. In several segmentation-classification-evaluation loops we determined the most important set of variables and the data layers providing them. Our main conclusions are (1) Multi-temporal data input greatly improve CAs of grassland and herbaceous vegetation classes in floodplains: user’s accuracies exceed 90%, and (2) the input data contributing most to these high CAs are NDVI layers from winter, spring and summer, and nDSM layers from winter and end of summer.