Within- and cross-regional crop classification for cool climate upland agriculture using AlphaEarth.

Accurate crop classification is essential for precision agriculture, policy making, and environmental monitoring. While Google AlphaEarth provides powerful analysis-ready geospatial data with high-dimensional embedded vectors, its applicability for crop classification is limited and cross-regional transferability remains untested. This study evaluated within- and cross-regional performance of crop classification using AlphaEarth across three cool climate agricultural regions: Northern Europe (Germany), Northern Japan (Hokkaido), and Northern US (Michigan). Within-region crop classification achieved high overall accuracy (91.4-99.1%) irrespective of regions and classifiers (support vector machine, random forest, and partial least squares). However, cross-regional transfer showed limited success. While models trained using Japanese dataset exhibited high accuracy (88.0-94.9%) in US dataset, reverse transfer (i.e. trained with US dataset and tested with Japanese dataset) and other combinations of transfer did not result in accurate classification. This study demonstrated (1) the utility of AlphaEarth as input data for regional-level crop classification, and (2) the limitations of AlphaEarth for cross-regional crop classification.