Machine Learning Procedures for Daily Interpolation of Rainfall in Navarre (Spain)

Abstract

Kriging is by far the most well known and widely used statistical method for interpolating data in spatial random fields. The main reason is that it provides the best linear unbiased predictor and it is an exact interpolator when normality is assumed. The robustness of this method allows small departures from normality, however, many meteorological, pollutant and environmental variables have extremely asymmetrical distributions and Kriging cannot be used. Machine learning techniques such as neural networks, random forest, and k-nearest neighbor can be used instead, because they do not require specific distributional assumptions. The drawback is that they do not take account of the spatial dependence, and for an optimal performance in spatial random fields more complex machine learning techniques could be considered. These techniques also require a relatively large amount of training data and they are computationally challenging to implement. For a reduced number of observations, we illustrate the performance of the aforementioned procedures using daily rainfall data of manual meteorological gauge stations in Navarre, where the only auxiliary variables available are the spatial coordinates and the altitude. The quality of the predictions is carefully checked through three versions of the relative root mean squared error (RRMSE). The conclusion is that when we cannot use Kriging, random forest and neural networks outperform k-nearest neighbor technique, and provide reliable predictions of rainfall daily data with scarce auxiliary information.