Tract-scale persistence of daytime land-surface temperature in Phoenix: LSTM-based screening beyond seasonality

Extreme heat is often assessed by peak intensity, yet many of its most damaging effects arise when extreme conditions persist across consecutive months, compounding exposure and limiting recovery. This study develops a neighborhood-scale persistence-screening framework for the Phoenix Metropolitan Statistical Area (Arizona, USA) using monthly MODIS/Terra land-surface temperature (LST) aggregated to 1,009 census tracts from January 2013 to November 2025. Tract thermal memory is characterized using lag-1 and lag-12 autocorrelation together with descriptors of seasonality, volatility, and residual variability. Predictability beyond seasonality is evaluated using a pooled Long Short-Term Memory (LSTM) model against seasonal-naïve, naïve persistence, and ridge autoregression benchmarks across 12-, 24-, and 36-month lookbacks. The LSTM outperforms the seasonal-naïve benchmark in 99.8% of tracts, with best performance at a 12-month lookback (mean RMSE = 1.984 °C; mean ΔRMSE = 0.34 °C), indicating an effective predictive memory of about one year. Historical persistence is generally short but becomes more evident in the late 2010s and early 2020s. A conservative 2030 screening map identifies an upper tail of persistence susceptibility concentrated in Central Phoenix, with secondary clusters in the West Valley and East/Southeast Valley. Overall, the results support tract-scale heat adaptation planning.