Automatic water level estimation from repeated crowd-based images of streams
Hi, I’m Wang Ze, a visiting PhD student in the CrowdWater project. My research interests include using machine learning methods for hydrological monitoring and modelling. In the CrowdWater project, I am combining citizen science and artificial intelligence to get more valuable hydrological information, especially water level data.
In my latest research, I moved from relying entirely on citizen scientists’ readings and evaluations to a deep learning technique as an auxiliary tool to automate water level class recognition from the images submitted by citizen scientists.
Specifically, a deep convolutional neural network model is introduced to retrieve water level classes from crowd-based images of stream banks. When estimating the current water level class based on an original image with a virtual staff gauge, the citizen scientists also upload an image showing the current conditions. Based on this pair of images, the model enables real-time checking of water level class observations by citizen scientists and can even be trained to derive higher resolution water level class time series with the help of votes by numerous participants in the CrowdWater game.
For future work, firstly, I will focus on the interpretation of the deep learning model mechanism. I want to explore if the computer thinks in the same way as us when it estimates the water level class. Furthermore, I aim to develop an automatic interaction mechanism between citizen scientists and artificial intelligence: the computer can instruct citizen scientists to upload the most informative images in real-time. With these informative images submitted by citizen scientists, the computer can be trained to be smarter.
Thanks to everyone for contributing valuable images and water level class estimates to the CrowdWater project. It’s actually your contribution that teaches artificial intelligence from knowing nothing at all to be smart! If you are interested in my research and want to learn more about how the model works, don’t hesitate to write to email@example.com.