Recognition of Aquatic Invasive Species Larvae Using Autoencoder-based Feature Averaging

EasyChair Preprint 8789

Abstract

The spread of invasive aquatic species disrupts ecological balance, damages natural resources, and adversely affects agricultural activity. There is a need for automated systems that can detect, track and classify invasive and non-invasive aquatic species using underwater videos without human supervision. In this paper, we intend to classify the larvae of invasive species like Zebra and Quagga mussels. These organisms are native to eastern Europe, but are invasive in United States waterways. It's important to identify invasive species at the larval stage when they are mobile in the water and before they have established a presence, to avoid infestations. Video-based underwater species classification has several challenges due to variation of illumination, angle of view and background noise. In the case of invasive larvae, there is added difficulty due to the microscopic size and small differences between aquatic species larvae. In video-based surveillance methods, each organism may have multiple video frames offering different views that show different angles, conditions, etc. Because there are multiple images per organism, we propose using image set based classification which can accurately classify invasive and non-invasive organisms based on sets of images. Image-set classification can often have higher accuracy even if single image classification accuracy is lower. Our proposed system classifies image-sets with a feature averaging pipeline that begins with an autoencoder to extract features from the images. These features are then averaged for each set. In our case, each set corresponds to a single organism. The final prediction is made by a classifier trained on the image set features. Our experiments show that feature averaging provides a significant improvement over other models of image classification, achieving more than 97% F1 score to predict invasive organisms on our video imaging data for a quagga mussel survey.

Keyphrases: Autoencoder, Classification, Feature Averaging, Quagga mussels, image set, invasive species

@booklet{EasyChair:8789,
  author    = {Shaif Chowdhury and Greg Hamerly},
  title     = {Recognition of Aquatic Invasive Species Larvae Using Autoencoder-based Feature Averaging},
  howpublished = {EasyChair Preprint 8789},
  year      = {EasyChair, 2022}}