Ensemble Machine Learning greatly improves ERA5 skills for wind energy applications

The skill of ERA5 has been assessed in relation to the prediction of the wind energy associated with 28 SYNOP stations located in Italy for a time span of 20 years(2001–2020).For comparison,a WRF-based high-resolution downscaling(3 km horizontally)was also produced for the same period.We found that simple predictions based on materialized past wind measures outperform the wind energy predictions from ERA5.This result can be ascribed to the particularly complex characteristics of the Italian territory.Motivated by this expected behavior,we have implemented a Quantile Random Forest(QRF)calibration which greatly alleviates the problems encountered in the ERA5 reanalysis dataset.This technique provides a calibrated ensemble prediction system for the wind speed at the station.Surprisingly,the calibrated ERA5 outperforms wind energy estimations from the high-resolution 3-km downscaling.Once properly calibrated,the high-resolution downscaling provides predictions very similar to the calibrated ERA5.Limiting our conclusions to the estimation of wind energy over a long time span(here 20 years),having at disposal a high-resolution wind-field dataset does not necessarily mean greater accuracy.A careful calibration of the original coarser wind-field dataset produces better results than the raw high-resolution dataset.

Comparison of sampling designs for calibrating digital soil maps at multiple depths

Digital soil mapping (DSM) aims to produce detailed maps of soil properties or soil classes to improve agricultural management and soil quality assessment. Optimized sampling design can reduce the substantial costs and efforts associated with sampling, profile description, and laboratory analysis. The purpose of this study was to compare common sampling designs for DSM, including grid sampling (GS), grid random sampling (GRS), stratified random sampling (StRS), and conditioned Latin hypercube sampling (cLHS). In an agricultural field (11 ha) in Quebec, Canada, a total of unique 118 locations were selected using each of the four sampling designs (45 locations each), and additional 30 sample locations were selected as an independent testing dataset (evaluation dataset). Soil visible near-infrared (Vis-NIR) spectra were collected in situ at the 148 locations (1 m depth), and soil cores were collected from a subset of 32 locations and subdivided at 10-cm depth intervals, totaling 251 samples. The Cubist model was used to elucidate the relationship between Vis-NIR spectra and soil properties (soil organic matter (SOM) and clay), which was then used to predict the soil properties at all 148 sample locations. Digital maps of soil properties at multiple depths for the entire field (148 sample locations) were prepared using a quantile random forest model to obtain complete model maps (CM-maps). Soil properties were also mapped using the samples from each of the 45 locations for each sampling design to obtain sampling design maps (SD-maps). The SD-maps were evaluated using the independent testing dataset (30 sample locations), and the spatial distribution and model uncertainty of each SD-map were compared with those of the corresponding CM-map. The spatial and feature space coverage were compared across the four sampling designs. The results showed that GS resulted in the most even spatial coverage, cLHS resulted in the best coverage of the feature space, and GS and cLHS resulted in similar prediction accuracies and spatial distributions of soil properties. The SOM content was underestimated using GRS, with large errors at 0–50 cm depth, due to some values not being captured by this sampling design, whereas larger errors for the deeper soil layers were produced using StRS. Predictions of SOM and clay contents had higher accuracy for topsoil (0–30 cm) than for deep subsoil (60–100 cm). It was concluded that the soil sampling designs with either good spatial coverage or feature space coverage can provide good accuracy in 3D DSM, but their performances may be different for different soil properties.