Fine Scale Surface Climate In Complex Terrain Using Machine Learning

Accurate and high spatial resolution (<100 m) surface climate information is crucial for process-based modelling in hydrology, ecology, agriculture, urban studies etc, especially in complex terrain landscapes where coarse grid resolution information (similar to 10 km) is inadequate to represent pronounced local variability. We used a machine learning-based workflow to predict high resolution (30 m) and sub-daily atmospheric variables fields of near-surface air temperature and humidity, and wind speed. The method used the Principal Component Analysis (PCA) decomposition applied on ground stations observations or Global Climate Model (GCM) residual error, in a sequence with bias correction and statistical models (Linear Regression-LR, Artificial Neural Network model-ANN and Empirical Quantile Mapping-EQM) to provide downscaling from large scale atmospheric conditions to complex terrain variability. The predictions described relationships of Principal Component (PC) scores dependent on GCM temporal variability on 6-hourly basis (with LR or ANN or EQM), and PC loadings dependent on topographic indexes to help providing horizontal sub-grid extrapolation. The methods were validated with a 1-year dataset from a dense weather stations network deployed in a complex terrain basin in tropical climate of Southeast Brazil. We present an exhaustive description of the PC modes daily/seasonal variability for each variable, and their spatial variability associated to the topography and thermal driven circulations. The predictions in general substantially improved accuracy when compared to GCM outputs, especially near the valley and in sheltered area where local effects are mandatories. Specially, ANN and EQM significantly improved the predictions at the variability of extreme events, such as the formation of strong cold air pooling or wetting in the valley.