Answers are grounded in this paper and cite their sources.
Landsat observations provide detailed land-surface information for investigating aerosol optical depth (AOD) estimation over heterogeneous surfaces. We evaluated the predictive relevance of neighborhood background and local variability and examined sensitivity to model design and input processing. A multisource dataset combined Aerosol Robotic Network (AERONET) observations with Landsat 8/9 predictors, meteorological variables, and ancillary surface data. XGBoost, a pure Transformer, and a hybrid Mamba–Transformer were compared under a common station-level split and 0% training masking. On 11,062 test records from 110 held-out stations, the hybrid model achieved a root mean square error (RMSE) of 0.0709. Paired station-bootstrap intervals for inter-model RMSE differences included zero. Retrained ablations increased the hybrid model’s RMSE to 0.1002 after removing nine neighborhood-mean features and to 0.1067 after removing five local standard-deviation features. Both types of neighborhood information therefore contributed predictive information under the evaluated setting. Additional diagnostics examined conditional errors, water-vapor substitution, target transformation, and compositing order. Estimates for sparsely sampled surface types and high-AOD conditions remained uncertain. A China-wide map illustrated model response on a nominal approximately 1 km grid. Independent spatial validation remains necessary to establish regional accuracy and effective AOD resolution.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: