A statewide tree inventory mapped 16.3 million non-forest crowns—and published where the model is weakest
WHAT THE EVIDENCE SAYS
A Scientific Data workflow combined 0.2-m multispectral aerial imagery, height data, terrestrial LiDAR reference surveys and deep learning to map 16.3 million predicted trees across Baden-Württemberg’s non-forest domain. The one-class tree detector achieved mAP@0.5 of 0.647, while the ten-class genus model achieved 0.382 and macro-F1 of 0.627. The release includes crown geometry, height, genus class, confidence, code and training annotations.
WHY IT MATTERS / OUR READ
The dataset is useful partly because it does not collapse uncertainty into a clean-looking map. Crown detection, genus assignment and geographic transfer have different confidence, giving planners a more honest basis for deciding which questions can be answered remotely and where local inventory still matters.
CAVEATS / UNCERTAINTY
Training and ground reference were concentrated in three cities, genus classes are coarse, understory and overlapping crowns can be missed, and predictions outside the main reference domain are extrapolations. It is an aerially visible crown inventory, not a complete stem census.
SOURCE
Scientific Data — data descriptor
View original evidence ↗
A German remote-sensing workflow mapped individual trees outside forests with genus labels, but detection was substantially stronger than taxonomic classification.
Read the Wild Lands assessment →