Towards Quantitative Spatial Models of Seabed Sediment Composition

Abstract

There is a need for fit-for-purpose maps for accurately depicting the types of seabed substrate and habitat and the properties of the seabed for the benefits of research, resource management, conservation and spatial planning. The aim of this study is to determine whether it is possible to predict substrate composition across a large area of seabed using legacy grain-size data and environmental predictors. The study area includes the North Sea up to approximately 58.44°N and the United Kingdom's parts of the English Channel and the Celtic Seas. The analysis combines outputs from hydrodynamic models as well as optical remote sensing data from satellite platforms and bathymetric variables, which are mainly derived from acoustic remote sensing. We build a statistical regression model to make quantitative predictions of sediment composition (fractions of mud, sand and gravel) using the random forest algorithm. The compositional data is analysed on the additive log-ratio scale. An independent test set indicates that approximately 66% and 71% of the variability of the two log-ratio variables are explained by the predictive models. A EUNIS substrate model, derived from the predicted sediment composition, achieved an overall accuracy of 83% and a kappa coefficient of 0.60. We demonstrate that it is feasible to spatially predict the seabed sediment composition across a large area of continental shelf in a repeatable and validated way. We also highlight the potential for further improvements to the method.

Fig 1. Study area.

Strong black line indicates study area. 1) Fladen Grounds; 2) Norwegian Trough; 3) Bristol Channel; 4) Dogger Bank; 5) Norfolk Banks; 6) Southern Bight; 7) English Channel; 8) German Bight 9. Friesian Islands

Fig 2. Additive log-ratio transformation.

Example of transformation of predictions from log-ratio space to M/S/G fractions for a single test set observation. Predictions are made in the log-ratio space and reverse transformed into M/S/G fractions. The observed value is shown in red, predicted value (conditional mean) shown in green. 95% prediction intervals represented by the grey rectangle.

Fig 3. Feature Importance Scores.

The importance of predictor features indicated by the random forest algorithm. The x-axis indicates the average decrease in node sum of squares when variable is used.

Fig 4. Observed and predicted values for 1,000 random observations from test set.

The left panels show predicted vs observed values for alr _m and alr _s along with prediction intervals. The diagonal line indicates y = x. The right panels show the observed values relative to the centre of prediction interval. Points are coloured according to whether the observed value is within the 95% prediction interval.

Fig 5. Partial dependence plots.

Bivariate partial dependency plots showing response of mud/sand/gravel fractions to water depth and mean current velocity, averaging out the effects of all other variables.

Fig 6. Spatial plots of predicted mud, sand and gravel content.

Fig 7. Spatial plots of classified predictions.

Fig 8. Spatial plots of alr _m and alr _s 95% prediction interval widths.