Quantitative water depth determination in large experimental timeseries through combining spectrophotometry and machine learning

Estuaries are important areas for ecosystems and societies, which are all affected by physical processes. Understanding these processes is crucial for ecosystem and urban resilience in a changing climate. Accordingly, we conducted physical scale experiments of estuaries with self-forming channels and bars in the Metronome (www.uu.nl/metronome), captured in overhead imagery from which we produced a large dataset of orthomosaics of more than 200,000 tidal cycles. We add the dye Brilliant Blue FCF to the water in the facility for conducting optical water depth measurements from these orthomosaics. However, such dye application is not straightforward and comes with several challenges related to color degradation, evaporation and spectral calibration. In this study we first developed a protocol implementing spectrophotometry to measure and maintain consistent dye concentrations in our lab, and to confirm the robustness of Brilliant Blue FCF. We then used the protocol to collect carefully designed training datasets of orthomosaics of various dye concentrations and various water depths, including 15 color bands from 5 nonlinearly related color spaces. We used these datasets to train random forest models to classify dry and wet cells, and for wet cells specifically, determine water depths through regression. We validated these models on datasets of burst and morphological imagery and observe comparable performances for dye concentrations between 1-5 mg/L (RMSE of 2.51-3.29 mm), whilst these models show notable differences in feature importance. Our workflows can be implemented for automatic data-driven analyses of our experimental dataset, and can be modified to other laboratories.