Image-based turbidity measurements

The objective of the SMARTLAGOON project is to develop a digital twin of the Mar Menor and its entire watershed. It is intended to give real-time insights of the state of the Mar Menor, and also forecast the risk of hypoxia in the short-term future. The digital twin integrates water quantity and quality models and real-time sensor data.

Regarding the real-time sensors, at SMARTLAGOON photrack is developing methods to measure turbidity via images. Water clarity can be used as in indicator for water quality assessment, its spatio-temporal monitoring is important for water management, and environmental protection of aquatic ecosystems.

There are available sensors which can measure turbidity, however they have to in contact with the water, one of the advantages of using image-based systems is that they are not in contact which makes their installation and maintenance easier.

We are investigating the potential of RGB images and machine learning techniques for water clarity monitoring of small water bodies. Specifically, we are Convolutional Neural Networks to learn and predict the described case of image to Colored Dissolved Organic Matter (CDOM) mapping. As the perception of water is quite difficult due to water reflection, water surface refraction, and day/night changes multiple modern, state of the art models were trained. We implemented a ResNet18 model, because these perception models are built for image classification and do not output a continuous real value. We crafted specific activation functions and connected them to a modified last layer.

This model was tested at Kristineholm, a 7 meters wide stream in Sweden, were the CDOM is continuously measured. The values ranged from 24 to 32ppb. We used 1,141 images taken during daylight for training and 258 were used for testing, the RMSE on the test sample was of 1.0079.

 

 

The results are very promising and as next step we are planning to install the system at the mouth of the Rambla of the Albujón and integrating it with the image-based discharge monitoring system.