Automating Surface Water Pollution Detection and Instance Segmentation

Water pollution stands as one of the most threats in the aquatic ecosystem. Aquatic environments are intricate, characterized by diverse compositions and dynamic environmental conditions. For instance, the manifestation of oil pollution in water exhibits a variety of patterns and structures, spanning from subtle sheens to more pronounced slicks. The precise segmentation and classification of this diverse range of oil patterns pose difficulties for automated systems. Hence, in this paper, we present automated detection, segmentation, and classification of various water pollution items such as oil spills, plastic wastes, dead bodies, acids, and eutrophication. Nevertheless, the absence of labeled images in this field has made training accurate pollution detection models challenging. The proposed model consists of Data Augmentation using a Generative Adversarial Network (GAN). For pollution segmentation a Mask R-CNN with feature pyramid network (FPN) based on RESNet is used. The model achieves Average Precision (mAP) of 86.4% for box segmentation and mAP of 85.7% for instance segmentation. The obtained mAP of the proposed model outperforms the state-of-art by 4.2% in terms of mAP.