Horizon-dependent level–shape decomposition for multihorizon ERA5-Land temperature-field forecasting in the Lanzhou–Xining region

Accurate monthly reanalysis temperature-field forecasts must recover both regional displacement and the spatial organization of anomalies, particularly when a region changes thermal state. We developed RegimeTransfer-DCSF (Deep–Causal Shape Fusion), a multihorizon framework that separates a forecast-causal regional level from median-centred spatial shape, learns short-lead shape through a compact fine–coarse attention hierarchy, and transfers shape responsibility from the deep branch to a stable anchor as lead increases. Using 12-month histories of 12 ERA5-Land variables, the model directly forecast future ERA5-Land 2-m air temperature at one-, three-, and six-month leads (h1, h3, and h6) over 39 county-level administrative units of the Lanzhou–Xining urban agglomeration. Within two rolling-origin validation folds, 114 horizon-specific cases yielded 114,114 county-domain grid-cell predictions. The selected DCSF fit achieved a pooled MAE of 1.131°C, a root-mean-square error of 1.494°C, and an R² of 0.9767, with 83.68% of absolute errors within 2°C; a fold-wise two-refit county ensemble retained a pooled MAE of 1.135°C. Averaged across the 114 county-area-weighted cases, the selected fit reduced MAE by 9.95% relative to seasonal climatology and by 20.35% relative to TimeFilter. On the complete 3,655-cell support, the four-refit DCSF mean ranked first in the ten-method rolling-validation benchmark at 1.174°C, with an across-refit standard deviation of 0.0024°C. During 25 h1 thermal-state changes, DCSF reduced area-weighted MAE by 0.301°C (19.79%) relative to TimeFilter, with a fold-stratified three-month moving-block bootstrap interval of 0.073–0.526°C. In a separate retrospective ERA5-Land analysis, annual mean warm-area fraction increased by 9.29 percentage points per decade during 2000–2024; the autocorrelation-aware three-year block interval was 2.72–15.52 percentage points per decade. Six of the 48 validation target months met the widespread-warm definition and formed highly connected warm footprints. Within this rolling-validation task, explicit level–shape responsibility delivered high absolute field accuracy and its clearest short-lead value during thermal-state changes.