Data homogeneity impact in tree species classification based on Sentinel-2 multitemporal data case study in central Sweden

Spatial information on forest composition is invaluable for achieving scientific, ecological, and management objectives and for monitoring multiple changes in forest ecosystems. The increased flow of optical satellite data provides new opportunities to improve tree species mapping. However, the accuracy of such maps is affected by training data, and in particular on the homogeneity of individual classes. Thus, we evaluated the effect of data homogeneity in tree species classification. We performed tree species classification by considering different ways to partition data into tree species classes. The class sets considered were (i) only mixed coniferous and mixed deciduous forest classes, (ii) single-species classes, (iii) single-species, mixed coniferous and mixed deciduous classes, and (iv) single-species, mixed coniferous and mixed deciduous classes and a true mixed class. Using data from the Swedish National Forest Inventory, we varied the threshold that defined dominating species. Tree species were classified for a study area in central Sweden using Sentinel-2 data and two classification approaches: Bayesian inference and random forest (RF). Images were selected by class separability and the most informative images based on variable selection with RF. The most informative images tended to be selected by both methods. However, in forests with tree species of similar spectral behaviour, image selection on the basis of class separability was found to be more reliable. More accurate classification results were achieved as the number of classes decreased and the threshold of plot purity increased. The Bayesian classification approach of only mixed coniferous and mixed deciduous classes gave the highest OA, always greater than 90%. When discriminating between pure plots of Birch (Betula spp.), Spruce (Picea abies), Scots pine (Pinus sylvestris) and Lodgepole pine (Pinus contorta), the best OA values were 84% for Bayesian and 80% for RF. In more complicated scenarios, RF resulted in higher overall accuracies (OA).