Quality-Gated Integration of Predicted Fields for Maize–Soybean Mapping with AlphaEarth Foundations

Timely field-level crop mapping remains difficult where current labels and reliable parcel boundaries are scarce. This study examined whether 2024 AlphaEarth Foundations (AEF) annual embeddings could support maize–soybean mapping through unsupervised clustering followed by label-assisted semantic assignment and whether Fields of the World (FTW) predictions could convert pixel clusters into reliable parcel-level maps. The 64-dimensional AEF features at 10 m resolution were clustered without crop labels in an agricultural region of northern Northeast China. After the cluster solution was fixed, the field observations were used for semantic assignment and spatially blocked assessment. At the 652 route-based reference points collected in this study, direct AEF clustering yielded F1-scores of 0.944 for maize and 0.942 for soybean, with a macro-F1 of 0.943 for binary maize–soybean retrieval. Unconditional parcel assignment increased spatial regularity but reduced agreement with the reference observations. A quality gate based on pixel support, within-field purity, and under-segmentation risk retained a macro-F1 of 0.936 while reducing small patches by 31.0%. Selective use of predicted fields reduced fragmentation while preserving most of the sample-based accuracy of direct AEF clustering, offering a way to control the loss of pixel-level crop information during parcel aggregation.