Turbidivision: a machine vision application for estimating turbidity from underwater images

Abstract

The measurement of turbidity serves as a key indicator of water quality and purity, crucial for informing decisions related to industrial, ecological, and public health applications. As existing processes require both additional expenses and steps to be taken during data collection relative to photography, we seek to generate accurate estimations of turbidity from underwater images. Such a process could give new insight to historical image datasets and provide an alternative to measuring turbidity when lower accuracy is acceptable, such as in citizen science and education applications. We used a two-step approach to a machine vision model, creating an image classification model trained on image data and their corresponding turbidity values recorded from a turbidimeter that is then used to generate continuous values through multiple linear regression. To create a robust model, we collected data for model training from a combination of in situ field sites and lab mesocosms across suspended sediment and colorimetric profiles, with and without a Secchi disk for visual standard, and binned images into 11 classes 0-55 Formazin Nephelometric Units (FNU). Our resulting classification model is highly accurate with 100% of predictions within one class of the expected class, and 84% of predictions matching the expected class. Regression results provide a continuous value that is accurate to ±0.7 FNU of true values below 2.5 FNU and ±33% between 2.5 and 55 FNU; values that are less accurate than conventional turbidimeters but comparable to field-based test kits frequently used in classroom and citizen science applications. To make the model widely accessible, we have implemented it as a free and open-source userfriendly web, computer, and Google Play application that enables anyone with a modern device to make use of the tool, the model, or our repository of training images for data collection or future model development.