TREE SPECIES RECOGNITION AT STANDS SCALE: VALIDITY TEST OF MULTI-TEXTURE EXTRACTED FROM MULTI- SEASONAL UAV-BASED IMAGERY

In order to evaluate the effectiveness of multi-type texture features of images of four seasons in pure stand tree species recognition, this research applied 5-band RedEdge-MX sensor to collect remote sensing data of four seasons and extracted eight texture features, including mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation, from 20 spectral bands. Maximum likelihood classification and random forest were adopted for the determination of the best window for texture extraction which resulted in the construction of optimal texture feature set in tree species recognition. Then, the performance of these texture feature sets along with their combinations in tree species recognition was analyzed. Experimental findings showed that the eight texture features of four seasonal data performed well in the recognition of pure stand tree species. Texture feature mean presented the highest performance (with overall accuracy of 88.8559%) and worst variance (84.8180%). The combination of eight texture features further improved the recognition accuracy of tree species (92.0599%) compared with single texture features. The recognition accuracy of tree species could be further improved by combining eight texture features with spectral band and digital surface model (92.7002%). Research showed that the application of multi-type texture features in typical seasons of spring, summer, autumn and winter fully captured the differences of various tree species in different bands and seasons, which could be applied to the effectively identify pure stand tree species in regular plots.