Case study on the optimal scaling factor for semantic segmentation of remote sensing image

Random scaling is a widely employed training technique to mitigate overfitting in semantic segmentation models, enabling models to handle images with varying scaling factors. However, existing deep learning frameworks for remote sensing images adopt a default scaling factor of 1.0 for single-scale testing. This raises the question of whether this scaling factor is optimal for single-scale testing. To address this question, this study investigates the optimal scaling factor across six models for three remote sensing datasets. The results show that the optimal scaling factor for the UAVid, LoveDA, and Potsdam was 0.75, 1.0, and 1.25, respectively.