Early-Season Mapping of Tobacco Cultivation Using a Stacked Generalization Ensemble Method: A Case Study in Changting County, China

Accurate and timely information on tobacco cultivation is essential for macro-level regulation, tobacco monopoly management, and farmer support in China. Remote sensing offers an efficient means for mapping tobacco, yet most studies focus on mid- to late-season identification, while early-season detection remains underexplored. Therefore, we integrated multi-source remote sensing data with the Google Earth Engine (GEE) platform to quantify separability between tobacco and competing crops, determined the optimal early-season phenological window and discriminative features, and developed an early-season tobacco mapping approach using a stacked generalization ensemble. The results show that January–February period provides the most effective early identification window for Changting County, with near-infrared (NIR) and water-sensitive features contributing most to class discrimination. The stacked ensemble achieved an overall accuracy of 87.17%, with producer and user accuracies of 85.97% and 92.28%, respectively, outperforming individual classifiers. At the township scale, the mapped tobacco area was strongly associated with tobacco purchase volumes (R2 = 0.819), supporting the reliability of the derived maps. Spatially, tobacco fields exhibited a pronounced corridor-like pattern, primarily influenced by proximity to water systems and terrain conditions. The proposed workflow provides tobacco management agencies with timely, spatially explicit information to support precise and intelligent tobacco production management.