Detection transformer-based approach for mapping trees outside forests on high resolution satellite imagery

High-resolution satellite imagery is a crucial data source for comprehensive trees outside forests (TOF) mapping at various spatial scales. Accurate identification of individual trees in satellite imagery remains challenging due to the heterogeneous nature of tree crowns, spectral similarities with other vegetation and the necessity to process large areas. The emergence of deep learning techniques, such as detection transformer models (DETR), offers new ways to analyse images more efficiently and accurately. In this study, we proposed an end-to-end approach for large-area TOF detection based on an established detection transformer architecture, specifically DETR with Improved deNoising anchOr boxes (DINO). We labelled 23,643 tree crowns with bounding boxes in 330 WorldView-3 image patches from the megacity of Bengaluru, India. Using this dataset, we trained and tested DINO for individual tree detection. In addition, we adopted a two-level tiling scheme and developed an R-tree-based Box Merging method to adapt to large images and remove redundant predictions more efficiently. Comparative analyses underscore the superior detection performance of DINO with a SWIN transformer as backbone, exhibiting an F1 score of 74% and an AP of 76%, surpassing other models such as Faster RCNN, YOLO, RetinaNet, DETR, Deformable-DETR, and DINO-Res50. For further validation we evaluated the proposed detection approach in two additional locations, Delhi and Shanghai, with varying image quality, achieving F1 scores of 87% and 73%, respectively. Our work advances remote sensing applications by providing a robust solution for large-scale TOF mapping and management. • DINO(SWIN backbone) outperformed Faster R-CNN, YOLO, and others in TOF detection. • Proposed a DINO-based end-to-end method with tiling and box merging for large-area TOF detection. • Tested on two diverse locations with varying image quality, showing strong applicability.