Accurate, crown-level maps of every shrub and tree are essential for managing rangelands, but AI models that create them from high-resolution imagery usually demand thousands of local training examples. We are testing whether a deep-learning model originally trained on African-savanna imagery can be reused in the semi-arid rangelands of New Mexico by adding only a handful of new labels. Using 0.5 m Maxar satellite scenes, the “out-of-the-box” model correctly identifies most isolated crowns yet systematically misses large mesquite and piñon–juniper clumps, reflecting structural differences between the two regions. Our next step is a quick, two-round fine-tuning cycle in which a human corrects a few hundred crowns (<1 % of the original label volume) and feeds those edits back into the network; this light update is expected to raise accuracy metrics sharply. Once refined, the model will generate crown maps from 2000 to the present, delivering the first multi-decade, individual-plant record of how grazing regimes, brush removal, and fire have steered woody cover at selected sites—and offering a scalable blueprint for monitoring dryland woody dynamics across the Southwest U.S. and beyond.