On the predictability of rainfall anomalies over the Southern Amazonia : a comparison between NMME and statistical models

Prediction of rainfall over the Amazonian rainforest during wet season is fundamental to assess the regional water and energy balance and global carbon-climate feedbacks. Previous observational analysis has identified some large-scale atmospheric dynamic and thermodynamics conditions that can influence the rainfall anomalies during the wet season. Based on these observed persistent conditions that started between June and August (JJA, dry season), we have developed and evaluated several statistical models to predict rainfall conditions during September to November (SON, early wet season) for the Southern Amazonia (5-15oS, 50-70oW). Multivariate Empirical Orthogonal Function (EOF) Analysis is applied to the following four fields during JJA from the ECMWF Reanalysis (ERA-Interim) spanning from year 1979 to 2015: geopotential height at 200 hPa, surface relative humidity, convective inhibition energy (CIN) index and convective available potential energy (CAPE), to filter out noise and highlight the most coherent spatial and temporal variations. The first 10 EOF modes are retained for inputs to the statistical models, accounting for at least 70% of the total variance in the predictor fields. Then the 12-fold cross-validation method is used to estimate the tuning parameters used in the regression algorithms. Ridge Regression and Lasso Regression are able to capture the spatial pattern and magnitude of rainfall anomalies. Compared with the seasonal prediction based on dynamical models, this statistical prediction system has better predictions than the seasonal predictions of the dynamic climate model. The statistical models show longer and more accurate predictive persistence of the rainfall anomalies. In addition, we use Logistic regression and Neural Networks to predict the categorical states of rainfall over the Southern Amazon by classifying the rainfall states into two categories, i.e., dry and wet. Our statistical models show overall better predictions of categorical rainfall states than the magnitudes of rainfall in our study region. The accuracy of the statistical prediction based on Neural Networks method can reach greater than 90%, which is much higher than the simple logistic regression method, indicating the non-linearity of the atmospheric processes. Both predictions of the magnitudes and states of rainfall anomalies can be combined to provide more accurate information. The models we have developed have broad implications on the future development of seasonal climate prodictions and can be used for real-time forecasts in the future.