Answers are grounded in this paper and cite their sources.
Dense georeferenced training data often combine local redundancy with rare responses concentrated in small areas. This study proposes a fixed-budget, long-tail-aware adaptive quadtree framework that integrates response stratification, severe-cluster protection, projected-coordinate partitioning, and observed spatial medoids. The analysis used 28,587 one-kilometer land-subsidence records from California’s San Joaquin Valley and repeated blocked validation over five grid origins and five folds per origin. At a 35% output budget, the adaptive representation achieved a severe-tail RMSE of 56.36 mm/yr, compared with 68.60 for unweighted full-data training, 56.59 for response-weighted full-data training, 55.88 for response-stratified random sampling, and 56.29 for response-stratified local pivotal sampling. Its non-tail RMSE and bias were 28.73 and −11.14 mm/yr, versus 23.51 and −6.30 mm/yr for unweighted full-data training. A coordinate-only k-nearest-neighbor reference achieved a tail RMSE of 29.42 mm/yr and hotspot IoU of 0.643 in the primary within-footprint setting. Across 30 prespecified sampling seeds, both response-matched baselines showed narrow severe-tail RMSE distributions that overlapped the deterministic quadtree result. Across 21 model seeds and the 5- and 10-km buffer analyses, the proposed framework consistently achieved lower severe-tail RMSE than unweighted full-data training. The proposed framework provides deterministic, spatially structured training data reduction under a prespecified output budget.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: