UAV sensing-based semantic image segmentation of litchi tree crown using deep learning

In this paper, we explore the capabilities of UAV sensing, and its photogrammetry product-orthomosaic by using convolutional neural network for semantic segmentation of litchi tree crown. We prepare an UAV-Litchi dataset consisting of 12 large orthomosaics having a total of 3000 image patches of size $256 \times 256$. The litchi tree crown is segmented by using the U-Net and smooth bending method. Finally, testing is performed over large aerial images of $3500 \times 3500$ size to provide a complete segmentation map of the crop region. The proposed model show mIoU score of 89 % and F1 score of 98 % with UAV-Litchi dataset, while mIoU score of 78 % and F1 score of 91 % with the benchmark Potsdam dataset.