CanopyYOLO: An improved YOLOv11-seg method with multi-source fusion and Instance Merging for Chinese fir canopy segmentation

The canopy is an essential component of tree-related information. However, obtaining accurate canopy information from remotely sensed images remains a significant challenge due to factors such as varying canopy sizes, overlapping trees and diverse forest environments. To solve this problem, an improved You Only Look Once version 11 segmentation (YOLOv11-seg) method for Chinese fir canopy segmentation named CanopyYOLO is proposed. The Backbone of CanopyYOLO is redesigned to be Multi-Source to fully learn the features of the canopy edge and height structure, enabling RGB and Canopy Height Model (CHM) data to be input into the Backbone simultaneously. The FeaturePyramidShared module is designed in the Backbone of CanopyYOLO to address the widespread issues of branch overlap and shadowing in fir canopies. The CaformerCGLU module and the Pinwheel-shaped Convolution are introduced to increase sensitivity to fine branch and needle contours and to improve detection of low-contrast, irregularly shaped, or small-scale forest canopies. To solve the issue of repeated detection of the same canopy across windows after chunking large Unmanned Aerial Vehicle (UAV) images, an Instance Merging method based on chunk overlap length and merging threshold is designed. Compared with the benchmark model YOLOv11-seg, CanopyYOLO achieves a m A P 50 - 9 5 M a s k of 62.6%, representing a 4% improvement. In terms of F1 score, CanopyYOLO achieves 92.7% in F 1 M a s k , which is 2.3% higher than the benchmark model. Additionally, the Instance Merging method results in a 10% improvement in individual-tree counting accuracy. The results indicate that CanopyYOLO plays a significant role in promoting large-scale forest surveys and precision forestry.