PTCNet - Tree crown detection model based on YOLOv3

PTCNet is a deep-learning urban tree crown detector, developed on YOLOv3 architecture with a Darknet-53 backbone.

Model Training
1. Model: YOLOv3 with a Darknet-53 backbone for 30 epochs.
2. Batch: 8
3. IOU: 0.5 (Overlap between the reference and detected tree crowns)
4. Non-maximum suppression: 0.1
5. Confidence score: 0.1

Based on the test dataset, the model had >75% precision and 62% recall of ground-truth data. However, based on other scenario test data, the performance was higher. It was developed for the city of Pullman located in the southeastern part of Washington. More information on the model can be found in our publication. The model will be continually updated to improve its generalization and localization, and the newer version will be uploaded.

The PTCNet model was trained using ArcGIS Pro 3.4, and the .dlpk files can be found here.

The high-resolution RGB image used is here

The training, testing, and scenario annotation shapefiles are here in the hand annotations folder.

The detected tree crown shapefiles are in the predicted annotation folder.

The entire tree crown, including training/testing annotations, predicted annotations, and other annotations, is available here.

See the distribution of the tree crowns in the study area: Video

If you find any of these resources useful in your work, cite our publication.