FIGURE 4.

Mean precision, recall, and f1 score (balance between false positives and false negatives) of convolutional neural networks detecting dugongs in test field (ULM) images (continuous line) versus test social media (WEB) images (dashed line) for all runs (R0, training with WEB images only; R2–R12, training with WEB images mixed with a number of ULM images equivalent to 2–12% of WEB images) (shading, SD of metrics evaluated on test ULM images). Values of all performance metrics are in Appendix S5