Downscaling soil hydrological mapping used to predict catchment hydrological response with random forests

Abstract

The Hydrology of Soil Types (HOST) classification scheme is based on national soil maps in the UK and has been developed by using expert knowledge to link soil hydrological indicators with conceptual models of surface and subsurface flow pathways through the soil profile. HOST has been optimised by regressing HOST class proportions against the Base Flow Index (BFI) and has been used to estimate catchment hydrological response and as input to land and environmental risk mapping. In this study, we performed spatial disaggregation of HOST class map unit polygons at a 1:250,000 scale in Scotland using the Random Forests (RF) classifier and selected environmental covariates to downscale HOST class mapping to 100 m grid resolution. Training datasets were developed using an area-weighted sampling scheme and by extracting HOST class information simultaneously by both single-class and multiple-class (complex) polygons by using the estimated within-polygon HOST class proportions as weights. The most-probable HOST-RF class maps along with surfaces of HOST class probability of occurrence were generated using 100 realisations from respective RF model runs each using a new training dataset. The performance of the disaggregated HOST-RF class maps for predicting catchment hydrological response was assessed by comparing BFI calculated using RF-predicted HOST class proportions with BFI calculated using flow data from gauges in 90 selected catchments. Results show that in this study using a larger number of training samples increases the risk of overfitting the RF model whereas lower sampling densities of 1 point per km2 were found to be sufficient and capable of promoting effective disaggregation of both single-class and complex polygons. Hydrological response predictions for the study catchments made with the disaggregated HOST-RF classes improved when class proportions were weighted using the generated probability surfaces and were very similar with the ones made with the original HOST class polygons, with small prediction improvements observed in the more low-lying and more slowly-responsive catchments when the HOST-RF classes were used. Overall, validation results indicate that the RF classifier was successful in predicting the spatial distribution of HOST classes in the study area, with the exception of HOST classes found within complex polygons areas dominated by soils with organic horizons of variable depth due to the difficulty of the algorithm to identify clear distinctions based on available covariate values. This study demonstrates the potential of applying digital soil mapping techniques to map soil hydrological functions by using an integrative hydropedological approach which links soils and hydrology at a landscape scale.