Crop multiclass classification for large-scale crop mapping using instance- and feature-domain integrated transfer learning

Efficient, accurate, and large-scale acquisition of planting patterns is crucial for food security. Although previous research has successfully achieved large-scale crop mapping using time-series features derived from remote sensing imagery, classification schemes integrating crop-specific rhythms and phenological knowledge remain insufficient under target-label-free conditions in the study area. This study proposed a Crop Multiclass Sampling Optimization Transfer Adaboost (CMSTA) method, which utilized a geoscience-informed instance-feature-domain adaptation approach to combine large-scale crop features with fine-scale phenological features. First, we used time-series Sentinel-1/2 imagery to identify high-quality simulated samples within the target domain through a combination of crop characteristics and spatial analysis. Next, the Phenological Feature Transformer Vector (PFTV), a feature-space transformation technique, was applied via transfer learning to augment diverse samples suitable for the fine-scale target domain. Finally, large-scale crop mapping was performed using a Random Forest classifier integrating both crop and phenological features. The experimental results demonstrated that: (1) Utilizing crop features, 476 credible samples were initially generated for the Yellow River Delta (YRD). These were subsequently expanded to 5,091 through phenological features and instance-feature-domain transfer. Integrating phenological and crop features notably enhanced model performance, increasing the learning curve score by 0.11; (2) The CMSTA method achieved superior crop mapping results within the YRD, with an Overall Accuracy (OA) and Weighted F1-score of 0.90. Compared with models using solely feature-domain transfer (OA: 0.44) and instance-domain transfer (OA: 0.69), CMSTA improved OA by 46% and 21%, respectively.; (3) When applied to large-scale mapping in the North China Plain (NCP), the crop mapping results were validated as reliable, achieving R 2 values greater than 0.6 against county-level statistical data. Specifically, wheat exhibited high consistency, achieving an R 2 value of 0.88. Overall, CMSTA facilitates scalable, highly accurate multi-crop classification with minimal samples, offering a robust solution for unified crop mapping across extensive regions. • Built a transferable crop-mapping framework using crop and phenology features • Increased available samples from zero to 5,091 via features transfer. • Improved local classification accuracy by about 25%, reaching 90% overall. • Demonstrated transferability for large-scale crop mapping across regions.