Using Explainable Machine Learning to Investigate the Controls of Vertical Accretion on the Mississippi River Deltaic Plain, Louisiana, USA.

The interplay between marsh vertical accretion, shallow subsidence, and surface elevation change is crucial for the sustainable management of the Louisiana deltaic plain. Yet, the mechanisms regarding marsh vertical accretion rates remain poorly understood due to the complex interaction of biological, physical, and hydrological processes across a range of spatial and temporal scales. Here, we leverage data from the largest coastal monitoring program in the world, the Louisiana Coastal Reference Monitoring System (CRMS), to study the drivers of vertical accretion across the deltaic plain. We split 245 CRMS stations into five different experimental groups that 1) consist of all the available data; 2) are divided into groups based on whether they are directly or indirectly nourished by fluvial sediment; 3) are categorized by hydrologic basin; and 4) are defined by marsh community type. For each experimental group we test the performance of an Extreme Gradient Boosting (XGBoost) model, a Least Absolute Shrinkage and Selection Operator (LASSO) regression model, and a multiple linear regression model to predict marsh vertical accretion rates. Preliminary results suggest that marsh accretion rates across the delta are too stochastic to obtain a general equation for accretion, with an R-squared value of ~ 0.40 achieved from repeated 5-fold cross-validation tests. We find that broadly segmenting CRMS sites by their hydrologic basins and degree of fluvial nourishment does not add predictive power compared to the baseline model trained on all the CRMS stations, making these groups inadequate to define the mechanisms of accretion. However, when subsetting the data by marsh community, our predictions of vertical accretion rates in freshwater marshes achieve an R-squared value of ~ 0.60. The most significant predictors were distance to the fluvial and oceanic sources and bulk density of accreted material, suggesting a simpler and more robust correlation with accretion. Importantly, we find that tides are commonly identified as a key variable in various groups for determining vertical accretion rates, highlighting the potential importance of tides even in a microtidal system. Future work includes refining this approach to develop a more generalized understanding of the drivers of sedimentation on this threatened delta.