LCRA predicts event-aligned canopy responses across unseen Qinling landscapes

Introduction Predicting event-aligned canopy variation in mountains requires integrating vegetation state, phenology, irregular satellite observations, and landform context. We asked whether a task-structured predictor could improve spatial transfer of short-horizon EVI responses across unseen Qinling landscapes. Methods We studied 80 response-blind matched sets in the Qinling Mountains, each comprising one mapped landform and two same-event non-landslide controls, yielding 991 set-events and 2,973 member responses across 32 spatial blocks. The landform-conditioned canopy response aggregator (LCRA) integrated dated MODIS EVI with climate-season, terrain, canopy-state summaries, and matched-member relations to predict subsequent-minus-pre-event (S-P) and subsequent-minus-impact (S-I) EVI. Models were evaluated with block-exclusive outer folds, and nominal composite-start-window results were compared with a complete 16-day support analysis. Results In the nominal analysis, the compact LCRA core ensemble achieved member R 2 = 0.8219 versus 0.7964 for the EVI-state-plus-season linear reference Δ R 2 = +0.0254; 95%; block × shared-onset interval 0.0037−0.0532; matched-contrast R 2 was comparable (0.3684 versus 0.3677). Requiring complete 16-day input support to end before the target window yielded member R 2 = 0.7810 in 80 sets and 0.7988 in all 90 eligible sets; in the latter, Δ R 2 versus linear was +0.0240 −0.0037 to 0.0547), while MAE decreased by 0.0041 EVI (0.0002–0.0087). Validation-based variance scaling reduced primary-task 95% interval width by 38.6% and brought empirical coverage from 99.4% to 92.9%, with improved probabilistic scores. Observed responses showed seasonal sign reversal: mean S-P was +0.0710 EVI in May–June and −0.1814 and −0.1849 in July–August and September. Pre-event canopy state provided a second ordered axis; calendar-standardized severity responses were negative but non-monotonic. The regional landform contrast centered near zero, with bidirectional conditional spatial variation. Discussion LCRA combines accurate spatial transfer with a phenology-centered account of event-aligned canopy trajectories. Complete temporal separation retained strong absolute accuracy, although relative R 2 gains remained uncertain. These observational contrasts describe canopy variation rather than causal effects of hot–dry exposure.