Regarding the issue of the persistent challenges in agricultural insurance, particularly the inaccuracies in plant census quantification and individual canopy assessment caused by canopy occlusion and heterogeneous orchard environments, which caused by the overlapping canopy and the complex orchard scenario, the B_SegNeXt method of semantic segmentation-edge detection dual-branch model was adopted to study how to extract tree crowns and accurately count for individual mango plants. A multi-scenario mango single tree canopy dataset was created to verify the generalization ability and segmentation ability of the model. Meanwhile, a multi-scale convolutional attention backbone network coupled with an edge detection module was designed, and an active fusion decoder was constructed to fuse edge features and semantic features. The regularization loss function was combined to alleviate the conflict of double-branch losses. Verification experiments were carried out using drone images from the main mango production areas in Hainan. The results show that by comparing the recognition effects of PSPNet, Deeplabv3, Deeplabv3 +, U-Net, SegNeXt, SegFormer, Segmenter and B_SegNeXt, the accuracy of B_SegNeXt has been significantly improved compared with the original model. It can be seen that this model can effectively improve the accuracy of canopy detection and counting of individual mango trees, providing technical support for agricultural insurance claims, government subsidy distribution, precise management of orchards, and planning of economic crop planting.