Regional Nonlinear Relationships Across the United States Between Drought and Tree-Ring Width Variability From a Neural Network

Abstract

Neural networks were previously applied to reconstruct climate indices from tree rings but showed mixed results in skill relative to more standard linear methods. A two-layer neural network is explored for purposes of reconstructing summertime self-calibrated Palmer Drought Severity Index (scPDSI) across the contiguous United States. Reconstructions using neural networks are more skillful than a linear approach at 75% of the gridboxes if evaluated by the coefficient of efficiency and at 54% when using the Pearson correlation coefficient. The increased reconstruction skill is related to the network capturing nonlinear growth-climate relationships. In the Southwest, in particular, a nonlinear response function captures a diminishing sensitivity of growth to moisture under wetter conditions, consistent with alleviation of moisture stress. These results indicate somewhat less-severe and more-stable incidences of drought over the past two centuries in the U.S. Southwest.