Comparative Analysis of Activation Functions in U-Net for Binary Water Segmentation using Sentinel-2 Imagery

ENG: The study examines how different activation functions influence the performance of a U-Net model applied to binary water-body segmentation in Sentinel-2 imagery. Using an identical training setup for each experiment, six nonlinearities—ReLU, Leaky ReLU, ELU, PReLU, Swish and RReLU—are individually substituted into the network while all other parameters remain fixed. Comparative evaluation on a held-out validation set reveals that Leaky ReLU provides the most balanced trade-off between precision and recall, making it the preferred choice for accurate water-mask generation. PReLU offers a similar but slightly lower performance, whereas ELU excels at capturing additional water pixels at the cost of more false positives. The findings highlight the importance of activation-function selection in remote-sensing segmentation tasks and suggest further exploration of advanced nonlinearities and larger, more diverse datasets to enhance generalization.