Individual Tree Crown Based Tree Species Classification from Very High-Resolution UAV Images

Kurzfassung

Very high-resolution optical imagery from un- manned aerial vehicles (UAVs) enables detailed tree species (TS) classification, surpassing the limitations of satellite imagery. In this paper, we propose an individual tree crown (ITC)- based loss and postprocessing approach, which can optimize classifying single tree species, designed to minimize the impact of overlapping canopies and dense mixed-species regions. Our methodology uses U-Net as the baseline model to integrate the proposed ITC-loss by treating each tree crown as a single unit to refine the predictions through a voting process. Furthermore, we apply a postprocessing method, merging the results with an ITC segmentation to refine the results. This reshapes the ITC and decides on a final species with a voting mechanism. Our results show that both approaches improve the mean intersection over union (mIoU) by 1.60 % for the best test, with a mIoU score for the baseline training of 42.03 % and 44.39 % after including our proposed methods.