Figure 1
From: Consumer-grade UAV imagery facilitates semantic segmentation of species-rich savanna tree layers
F1-Scores of the models on specific test data sets. The three panels on the left compare the performance of U-Net, FC-DenseNet and DeepLabv3 + trained on orthomosaics of Blocks 1, 2, and 3 with different input tile dimensions (256, 512, and 1024 pixel width and height) and tested on plots 4, 6, and 7. The rightmost panel shows F1-scores of DeepLabv3 + trained with 512 × 512 pixel tiles from 15 ha of Blocks 1, 2 and 3 recorded in 2021 (a) tested on the remaining 3 ha of the same blocks, and (b) tested on the 2022 images of these 3 ha. Colours indicate input tile dimensions.
