Neural Networks Can Infer Geomorphic Model Parameters From Topography, and We Might Be Able To Figure Out How!

Jo Martin, CU Boulder Boulder Colorado, United States. jo.martin@colorado.edu

From the classic U-shaped glacial valley to the convex soil-mantled hillslope, geomorphic processes leave clear signatures on the landscapes they create. However, it has been challenging to develop topographic metrics that can be used to extract process parameters. While researchers have gained significant insights into geomorphic processes through metrics like mean local relief, channel steepness, and ridgetop curvature, it is still difficult to make quantitative predictions about processes from quantitative topographic measurements.

Prior modeling work has found that in 2D models that combine stream incision with diffusive hillslope processes, valley spacing is strongly controlled by the relative rates of advective and diffusive processes (Perron et al. 2008). In this work we train a simple convolutional neural network to predict the ratio of the coefficient of stream erosion (K) and coefficient of diffusion (D) used to generate the model topography. Across a test set of 1800 model runs with different K and D values, the network had a normalized root mean square error of 0.03, showing that convolutional neural networks have significant promise for extracting complex and geomorphically meaningful topographic signatures.

In this work we focus on interpreting the neural network to try to help explain what it is calculating in a theoretically grounded way. The output of activation maximization, Fourier analysis, neuron ablation, and other interpretability techniques are complicated, but might imply that the network is detecting patterns that are geographically meaningful. This poster will present these interpretability results.

2025 CSDMS meeting-069

Neural Networks Can Infer Geomorphic Model Parameters From Topography, and We Might Be Able To Figure Out How!