Palm Tree Inventory From Aerial Images Using Retinanet

Systematic tree inventories based on field surveys are resource-intensive. Remote sensing data has shown added value herein, and current high-resolution images open the opportunity to apply novel methods such as deep learning neural networks achieving even higher accuracy on tree detection. In this study, an optimized palm tree inventory model is proposed based on the RetinaNet object detector and high-resolution RGB images. The aim was to classify and locate palm trees in different scenes, with a different appearance, and with varied ages. We used orthophoto imagery available from the municipality of Elche (Spain) with a spatial resolution of 25 cm. Date palms and canary palms, grown in the region, were inventoried. The detection model for palm tree inventory achieved a precision of 89.3% (mAP@IoU = 0. 50) on the validation dataset and 76.9% (mAP@ IoU = 0.50) on the test dataset. These results demonstrate the promising outcomes of applying deep learning approaches from computer vision to remote sensing image analysis.