Spatial downscaling of precipitation using adaptable random forests

Abstract

This paper introduces Prec‐DWARF (Precipitation Downscaling With Adaptable Random Forests), a novel machine‐learning based method for statistical downscaling of precipitation. Prec‐DWARF sets up a nonlinear relationship between precipitation at fine resolution and covariates at coarse/fine resolution, based on the advanced binary tree method known as Random Forests (RF). In addition to a single RF, we also consider a more advanced implementation based on two independent RFs which yield better results for extreme precipitation. Hourly gauge‐radar precipitation data at 0.125° from NLDAS‐2 are used to conduct synthetic experiments with different spatial resolutions (0.25°, 0.5°, and 1°). Quantitative evaluation of these experiments demonstrates that Prec‐DWARF consistently outperforms the baseline (i.e., bilinear interpolation in this case) and can reasonably reproduce the spatial and temporal patterns, occurrence and distribution of observed precipitation fields. However, Prec‐DWARF with a single RF significantly underestimates precipitation extremes and often cannot correctly recover the fine‐scale spatial structure, especially for the 1° experiments. Prec‐DWARF with a double RF exhibits improvement in the simulation of extreme precipitation as well as its spatial and temporal structures, but variogram analyses show that the spatial and temporal variability of the downscaled fields are still strongly underestimated. Covariate importance analysis shows that the most important predictors for the downscaling are the coarse‐scale precipitation values over adjacent grid cells as well as the distance to the closest dry grid cell (i.e., the dry drift). The encouraging results demonstrate the potential of Prec‐DWARF and machine‐learning based techniques in general for the statistical downscaling of precipitation.

Spatial Patterns of Water Age: Using Young Water Fractions to Improve the Characterization of Transit Times in Contrasting Catchments

“…Random forests are robust against overfitting (i.e., explaining the training data instead of finding general patterns) even in case of small data sets (Breiman, ). Applications of random forests in hydrology include empirical simulation of monthly streamflow (Shortridge et al, ), downscaling of precipitation data (He et al, ), and evaluation of flood hazard risk (Wang et al, ). In this study, we used the random forest implementation in the R statistical software (package “randomForest”) to quantify the prediction strength, expressed as variable importance, of the following predictors for F yw , α , and β : mean catchment slope, median surface flow path length, catchment area, percentage of agricultural land, forest and urban areas in 2012, soil fractions of sand, silt, and clay, and the mean values of annual precipitation, annual PET, annual runoff coefficient (determined from modeled discharge), and daily baseflow index (determined from event separation of modeled discharge) between 2013 and 2015.…”

Water Resources Assessment of China's Transboundary River Basins Using a Machine Learning Approach

“…Ensemble learning has been applied in many disciplines and areas and displays advantages over traditional algorithms (Deville et al, ; X. He et al, ; Stevens et al, ). Among the ensemble learning methods, RF and GB are widely used.…”

“…Machine learning explores the relation between the response and its relevant predictors using one or multiple algorithms, with no need to consider the explicit mathematical form of the model (Elith et al, ; X. He et al, ). Examples of machine learning algorithms include nearest neighbor (Cover & Hart, ), naïve Bayes (Jensen, ; Lewis, ), decision trees (Breiman et al, ), support vector machines (SVMs; Vapnik, ), and artificial neural networks (ANNs) (Hopfield, ).…”

Leveraging machine learning for predicting flash flood damage in the Southeast US

“…This method is suitable for both regression and classification problems. Due to randomized and decorrelated features of RF, it is able to build the connection between the input and output variables when their relationship is very complex and nonlinear ( He et al 2016, Hong et al 2016.…”