Predicting bankfull channel dimensions through Stepwise Multiple Linear Regression and Random Forest in intermittent Mediterranean streams.

Hydrological and soil erosion models are often used to assess the impacts of global change and potential adaptation strategies on flood risks and sediment transport. These hydrology and sediment transport models require channel dimensions as input to quantify flood frequency, runoff, flow velocity, sediment detachment and deposition processes. Especially for large-scale applications, channel dimensions (width and depth) are difficult to obtain. Therefore, simple empirical relations have been developed, relating channel dimensions with catchment area or bankfull discharge, disregarding other important factors affecting these dimensions.   

Here we present an advanced combined methodology to obtain reliable estimates of channel dimensions for the large Mediterranean Segura catchment (16,000 km²), based on linear statistical regression and machine learning techniques. First, a training dataset of channel dimensions (width and depth) was prepared using a LiDAR high resolution digital elevation model (2 m resolution) and aerial photos (50 cm resolution) for 151 channel segments across four representative large sub-catchments. For each channel segment, 30 variables characterising the upstream catchment were obtained from available spatial data sources (e.g. soil type, slope, annual precipitation). The obtained training dataset was used in a combination of Stepwise Multiple Linear Regression and Random Forest to predict channel width and depth. Best results were obtained with the RF model using the variables selected through the stepwise MLR process, as RF models composed only by these MLR predictor variables showed nodes with more purity rather than RF formed by the complete set of independent variables. Most important variables for prediction of channel width were Calcareous lithology, mean annual temperature, extreme precipitation, and alluvial soils. For channel depth, the most important variables were extreme precipitation, channel slope, and mean annual temperature. Model validation indicated good results for prediction of channel width (R² 0.75) and depth (R² 0.66). These results provide further insights into the factors affecting channel dimensions, and seems to be a promising approach to obtain channel dimensions for hydrological and sediment transport modelling in large catchments.

We acknowledge funding for the XTREME project from the Spanish Ministry of Science and Innovation and ‘Agencia Estatal de Investigación’ (PID2019-109381RB-I00/AEI/10.13039/501100011033), and for the LandEX project (PCI2024-153454) financed by the European Commission, Ministry of Science, Innovation and Universities and the Spanish Research Agency (AEI 10.13039/501100011033/EU) in the framework of the European Water4All Partnership 101060874. A. Jodar-Abellan (JDC2022-049314-I) and J.P.C. Eekhout (IJC2020-044636-I) acknowledge financial support from the Ministry of Science, Innovation and Universities for the Juan de la Cierva postdoctoral grants.