Diagram from the paper.
Catchment characteristics that might impact dissolved organic carbon (DOC) export - including topography, geology, and vegetation - were used to identify four clusters of sites with similar behaviors. Within each cluster, the most important characteristics that predict DOC export were identified with an evolutionary algorithm. Key findings separate low- and high-elevations (e.g., clusters 1-2 vs. 3-4), sediment and rooting depth, and vegetation type. Credit: Underwood et al. [2023], Figure 7
Editors’ Highlights are summaries of recent papers by AGU’s journal editors.
Source: Water Resources Research

In catchment science, most studies rely upon deductive methods of science, slowly and methodically assembling ever-more-specific process understanding. The goal, then, is the assembly of these facts to describe complex systems by combining processes in a bottom-up approach to replicate emergent properties and patterns.

In contrast, Underwood et al. [2023] eschew this tactic for an inductive approach, leveraging the strength of machine-learning to identify patterns and explanatory variables. In this application, an evolutionary algorithm is used to sift through a network of hundreds of catchments and nearly 2.7 million dissolved organic carbon (DOC) observations to identify which of 54 possible attributes explain annual DOC export. Critically, the predictions are not taken as a statistical ‘black box’, but instead are integrated with domain knowledge in a discussion section titled From Pattern to Process.

This approach is a strong template for researchers, where the results of the inductive, machine-learning approach are taken as hypotheses to be tested, either by comparison to existing knowledge (e.g., the authors’ mechanistic explanation of high DOC loads being correlated to thicker overburden and deeper rooting depth) or as an opportunity for subsequent, deductive investigation (e.g., unexplained patterns of high DOC efflux in the midwestern United States). Ultimately, this is a model of responsible application of machine learning to efficiently explore complex datasets, identify alignment with pre-existing understanding, and to inspire hypothesis generation to explain anomalies in the data set.

Citation: Underwood, K. L., Rizzo, D. M., Hanley, J. P., Sterle, G., Harpold, A., Adler, T., et al. (2023). Machine-learning reveals equifinality in drivers of stream DOC concentration at continental scales. Water Resources Research, 59, e2021WR030551. https://doi.org/10.1029/2021WR030551

—Adam S. Ward, Associate Editor, Water Resources Research

Text © 2023. The authors. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.