Annotated tree crown bounding boxes in urban/rural environment

This dataset contains 4782 tree crown annotations intended for use in training or evaluating tree crown detection models, such as those developed under the MillionTrees project. Two distinct sets are included:

Dataset 1 (jelled_annotations_finetuning_DF): 490 PNG image tiles (640×640 pixels) at ~0.094 m/pixel resolution, each annotated with bounding boxes for a total of 3,408 individual trees. While the data is not georeferenced, filenames encode the coordinates of the upper-left corner of each tile.

Dataset 2 (jelled_annotations_FW_DF): 98 merged PNG tiles, each formed by horizontally stitching three original tiles (from Dataset 1, at ~0.094 m/pixel resolution) . These contain annotations for 1,374 solitary trees. No trees in clusters or dense forest canopies are annotated in this set. Merged tile filenames do not contain coordinates, and this dataset does not overlap with Dataset 1.

The dataset is intended to support training and evaluation of object detection models like DeepForest. All annotations are in CSV format with the bounding boxes as: "image_path,xmin,ymin,xmax,ymax,label,width,height". Both datasets were annotated using Label Studio.