Optimize Plankton Image Classification
Abstract
To understand the impact of climate change on the marine life; researchers have been monitoring and investigating its effect on the food supply chain, among many other things. Plankton form the base of the marine food network and they are key to the sustainability of marine ecosystem. Early life stages are a major bottleneck for marine fish population, and thus understanding the drivers of larval growth and survival is key research topic in fisheries science and marine ecology. Thus, obtaining reliable estimates of the prey fields available (both in terms of abundance, diversity and size) to the fish larvae in the wild can inform about the starvation potential in the larvae, an important mortality source in this stage. In this research work contribute to the on going effort to map zooplankton population and to classify its types based on images captured by a FlowCam instrument for the past decade(2013-2019) and on ZooScan images taken using underwater vision profiler. State of the art image processing and machine learning algorithms had been applied to solve the plankton classification problem. However, with current advances in machine learning and computation power, the task can be optimized the accuracy and improve the results to provide more insights and information about the health and growth of these populations. Some have very high accuracy, while others very low or are being mis-classified - meaning there is still a need for human expertise to verify classification. Even experts can make mistakes, but by optimizing the state-of-the-art methods that are in use classification can happen faster and more precise, supporting research on the fate of marine fish populations under a changing climate. In this project we were able to get a training accuracy of 98\% and F1 score of 88\% using optimized MobileNet, which is comparative to the state-of-the-art.