Classification of urban tree species using multi-features derived from four-season RedEdge-MX data

Abstract

In the remote sensing discrimination of tree species, the use of multi-temporal images may considerably improve the accuracy of recognising tree species. This study aimed to explore the effectiveness of the RedEdge-MX sensor, which was used to capture images during the key node period of tree growth, for driving the identification of tree species. The spectral band, texture and digital surface model (DSM) features of these data and their combinations were used as data sets, and 32 tree species were classified using maximum likelihood classification and random forest. The results demonstrated that the image imaging period considerably influenced the recognition of tree species. The recognition accuracy of tree flowering and leafing period data was the highest (52.98%, 86.66% and 86.90% based on spectral, texture and spectral + texture + DSM features, respectively), whereas that of the leafy period data was the lowest (34.32%, 82.39% and 82.81%). The classification accuracy significantly improved (72.76%, 91.51% and 92.16%) when multi-temporal data were combined. Moreover, the texture extraction window significantly affected the classification accuracy. The accuracy was low (70.47%, four seasons) when the minimum window was used, and this rate increased to 91.52% when the appropriate window was used but declined when an excessively large window was used. Furthermore, the classification accuracy obtained using texture features was higher than that of the spectral bands and DSMs were used. Their combination optimised the accuracy of tree species classification (92.16%). These results show that the capture of key node images in tree growth, the selection of appropriate texture extraction windows, and the use of multiple types of features can drive the effective recognition of urban greening tree species.