NOAA/WDS Paleoclimatology - Ni et al. 2002 Southwestern USA Linear Regression and Neural Network Precipitation Reconstructions

A 1000 year reconstruction of cool-season (November–April) precipitation was developed for each climate division in Arizona and New Mexico from a network of 19 tree-ring chronologies in the southwestern USA. Linear regression (LR) and artificial neural network (NN) models were used to identify the cool-season precipitation signal in tree rings. Using 1931–88 records, the stepwise LR model was cross-validated with a leave-one-out procedure and the NN was validated with a bootstrap technique. The final models were also independently validated using the 1896–1930 precipitation data. In most of the climate divisions, both techniques can successfully reconstruct dry and normal years, and the NN seems to capture large precipitation events and more variability better than the LR. In the 1000 year reconstructions the NN also produces more distinctive wet events and more variability, whereas the LR produces more distinctive dry events. The 1000 year reconstructed precipitation from the two models shows several sustained dry and wet periods comparable to the 1950s drought (e.g. 16th century mega drought) and to the post-1976 wet period (e.g. 1330s, 1610s). The impact of extreme periods on the environment may be stronger during sudden reversals from dry to wet, which were not uncommon throughout the millennium, such as the 1610s wet interval that followed the 16th century mega drought. The instrumental records suggest that strong dry to wet precipitation reversals in the past 1000 years might be linked to strong shifts from cold to warm El Niño-southern oscillation events and from a negative to positive Pacific decadal oscillation.

针对美国亚利桑那州与新墨西哥州的各气候分区，研究人员基于美国西南部19个树轮年表网络，重建得到1000年尺度的冷季（11月至次年4月）降水序列。研究采用线性回归（linear regression, LR）与人工神经网络（artificial neural network, NN）模型，以识别树轮中的冷季降水信号。基于1931–1988年的器测记录，逐步线性回归模型通过留一法完成交叉验证，人工神经网络模型则采用bootstrap（自举）技术开展验证。上述最终模型还采用1896–1930年的降水数据进行了独立验证。在多数气候分区中，两种模型均可成功重建干旱年与正常年份的降水序列；相较线性回归模型，人工神经网络模型更能捕捉极端降水事件与更显著的降水变异性。在1000年的重建序列中，人工神经网络模型同样能生成更具特征性的湿润事件与更显著的变异性，而线性回归模型则能更清晰地反映干旱事件特征。两种模型重建的1000年降水序列均呈现出若干持续的干湿时段，可与1950年代干旱（例如16世纪特大干旱）以及1976年后的湿润期（例如1330年代、1610年代）相类比。极端气候时段对环境的影响，在干旱向湿润的突变转换期间往往更为显著；这类转换在千年尺度的记录中并不鲜见，例如紧随16世纪特大干旱之后的1610年代湿润时段。器测记录显示，过去1000年间发生的强干旱-湿润降水转换事件，可能与厄尔尼诺-南方涛动（El Niño-Southern Oscillation, ENSO）从冷位相向暖位相的显著偏移，以及太平洋年代际振荡（Pacific Decadal Oscillation, PDO）从负位相向正位相的转变存在关联。