Deep Reinforcement Learning for Macro-Scale Dissolved Organic Carbon Predictions from Multi-Modal Data

Dissolved organic carbon (DOC) is a key water quality constituent providing a key energy source for freshwater ecosystems, altering light regimes in freshwaters, and controlling nutrient cycling dynamics. However, it is underrepresented in freshwater data collections and databases when compared to other datasets on nutrients, algae concentration, sediment and other key water quality parameters. To fill this data-gap we propose estimating DOC concentrations with remote sensing and deep learning.  We aim to construct a generalized model capable of predicting DOC across freshwater estuaries, lakes, and rivers leveraging the Landsat archive and our dataset that pairs Landsat reflectance data with in-situ DOC collections.   To accomplish our goal, we introduce an innovative approach employing deep reinforcement learning (DRL) for accurate predictions of DOC concentrations from remote sensing data across the contiguous US. DOC varies spatial-temporally, and are highly responsive to terrestrial inputs, biological activity, land-use, human activity, and water body metabolism. Because of these highly varied controls on DOC dynamics, we complement our direct remote sensing data with additional inputs to produce reliable national data.  Although model performance approximately matches current models with an MAE of 1.93 mg/L overall, our approach provides specific error rates for types of water bodies in ecoregions.  This work forms the foundation of future work exploring controls on DOC variation in waterbodies of the USA, with an unbiased national data source.  

Primary Presenter: Daniel Dominguez, Colorado State University (Daniel.Dominguez@colostate.edu)

Authors:

Daniel Dominguez, Colorado State University  (Daniel.Dominguez@colostate.edu)

Bethel Steele, Colorado State University  (b.steele@colostate.edu)

Oktay Karakus, Colorado State University  (karakuso@cardiff.ac.uk)

Matthew Ross, Colorado State University  (matt.ross@colostate.edu)