Summer · monsoon & flood
The North American Monsoon recharges groundwater but also delivers intense, localized storms. This is where we project flash-flood and extreme-precipitation risk.
Why summer matters
The monsoon supplies roughly a third of Arizona's annual precipitation, some of it recharging shallow groundwater. But on sun-hardened soil and steep terrain, 1 to 2 inches of rain in an hour can flood open pits, threaten tailings facilities, and mobilize acidic runoff. Monsoon years are trending toward the extremes, either very dry or very wet.
Early and late summer REOF, then a dCNN
We split the monsoon into early summer (June to July) and late summer (August to September), which are governed by different drivers, and compute rotated empirical orthogonal function (REOF) patterns for each. Early summer is tied to ENSO and Pacific decadal variability; late summer decouples from Pacific forcing, with Indian monsoon convection emerging as a driver. Each season's REOF patterns are regressed onto 500 hPa height, sea surface temperature, and 200 hPa wind. A dilated convolutional neural network (dCNN) then trains on those historical patterns to reproduce daily precipitation, and runs on Seamless System for Prediction and Earth System Research (SPEAR) future fields to downscale to high-resolution daily rainfall. Contingency, if the dCNN underperforms or future patterns diverge: tree-based learners (random forests, gradient boosting), an multilayer perceptron (MLP) on the leading components, or an long short-term memory (LSTM), all using the same REOF-reduced predictors.
Monsoon downscaling and forecasts
SPEAR and the extended CAM6 runs provide the future large-scale fields the downscaling runs on. SPEAR's 30-member large ensemble captures how the monsoon ridge may shift by 2050, while CAM6 adds realistic decade-long swings driven by Pacific and Atlantic ocean variability. Season to decadal (S2D) model runs add 1 to 5 year initialized outlooks for summer precipitation, giving the mid-term horizon the mining sector plans against. The downscaled product is released openly on the Data and Code page. Models used here: SPEAR, CAM6, S2D.
Note on resolution
The downscaling produces high-resolution daily precipitation fields, but these are statistically downscaled outputs, not validated point forecasts.