Enhancing Water Surface Detection with Domain-specific Data Augmentations

Detecting water surfaces in various environmental conditions is crucial for applications such as autonomous navigation and environmental monitoring. The variability in water appearance due to factors like lighting, reflections, and sur face disturbances makes this task challenging. We address these challenges by proposing domain-specific data augmentation techniques using the Albumentations library, including varying brightness, adding reflections, simulating waves, and changing color tones. We implement these augmentations in the data loader and train a DeepLabv3 model with the augmented images. Our approach is validated by comparing the performance of the augmented model with a baseline model on diverse water conditions. Experimental results demonstrate that our augmentations significantly improve the model’s generalization and robustness in water surface detection tasks.