RESEARCH ON THE EFFECT OF ACTIVATION FUNCTIONS ON THE ACCURACY OF WATER BODY SEGMENTATION FROM SATELLITE IMAGES

A comparative analysis of various activation functions (ReLU, Leaky ReLU, ELU, PReLU, Swish, RReLU) was conducted within a U-Net neural network architecture for the purpose of binary segmentation of water bodies in Sentinel-2 satellite imagery. The experimental design employed a fixed number of training epochs and a consistent set of images, thereby enabling an objective assessment of the results. The findings demonstrate that the selection of the activation function substantially influences the effectiveness of neural network-based segmentation. Among the evaluated functions, Leaky ReLU produced the highest F1 score, as well as superior Precision and IoU values. In contrast, while the ELU function exhibited the highest Recall, its markedly lower Precision indicates a tendency toward over-segmentation. Based on these observations, Leaky ReLU is recommended as the optimal choice for water-body segmentation in satellite images.