Dodge or Disinfect? Classifying Algae in Small-Scale Water Bodies via Low-Cost Deep Neural Networks

In recent years, small water bodies like lakes, ponds, and artificial rivers in cities have been widely seen the rapid growth of water plants, such as algae. As a result, unmanned water vehicles are deployed to clean up the water pollution. Most of these vehicles are manually controlled to determine whether the water plants are algae or landscape. A computer vision-based auto system will be helpful to alleviate the overload labor inefficiency. In this work, we conducted a series of experiments using transfer learning with three SOTA deep neural networks (VGG16, ResNet50, and Vision Transformer), to examine their classification performances on two types of identification standards, three water plant species and plant species in the near or far distance. In addition, we also examined the performance with data augmentation. Our result suggests that: (1) Vision Transformer overperforms in both classification tasks; (2) plant species standard works the best that leaves the result for augmentation in different times still varies depending on choices of other factors. Thus, our work contributes to three aspects: 1) Provide an inhouse dataset collected in the real-world small water bodies. 2) Provide model selection guidance by comparing three SOTA deep networks performances. 3) Apply vision transformer at the first time to the algae classification task and the benchmark enables the potential to extend the backbone model to downstream tasks, such as algae detection.