Improving extreme weather generation and spatial variability of weather generators based on resampling using the Direct Sampling algorithm
Weather generators are tools for generating synthetic data of weather variables. Those simulations need to be statistically equivalent to observed data and to reflect the weather persistence, variability and reproduce the spatio-temporal dynamics and correlation structures among the different meteorological variables. Resampling methodologies are powerful and simpler strategies for generating synthetic weather data that capture the observed statistical properties of weather. However, although weather generators using resampling can generate new time series, the spatial weather fields they produce are replicas or calibrated versions of the observed historical data. We propose in this work to overcome the spatial variability and the out-sample extreme event generation limitations by using the Direct Sampling (DS) algorithm as a postprocessing step on the weather generator outputs. The DS considers each weather field produced by the weather generator as a training image and a simulation grid of the same size for storing the simulation. Further, a set of random conditioning data points through the time and space are used to keep the original statistical properties of the original weather. Also, as weather generators based on resampling cannot generate weather fields correlated with hypothetical extreme scenarios unseen in the original data, we are proposing two ways of Generating of precipitation fields with extreme events: 1) quantile mappings applied on the conditional simulations from Direct Sampling correlated with return precipitation levels for a given return period, and 2) Direct sampling on a target weather field conditioned on a return precipitation level map for a given return period. We used a set of control points which can be randomly, or arbitrarily defined to select the areas where a modified version of the Direct Sampling algorithm will generate high precipitation values, coherently, using the return precipitation level as a guide. We conducted a series of experiments using several metrics to show that the DS outputs preserve the structural and connectivity patterns, and statistics observed in simulations. Also, we show that it is possible to generate hypothetic scenarios, keeping some connectivity, structural and statistical information from the original weather fields.