Lightweight YOLOv11 tree species recognition method based on bark texture

Objective To address the issue of existing tree species identification methods being difficult to deploy on mobile or edge devices with extremely limited hardware due to high computational complexity under varying lighting conditions, this study proposes a lightweight tree species identification method based on bark texture.

Method This research improves YOLOv11 to construct the YOLOv11-SWER model. First, the lightweight feature extraction network StarNet was introduced as the backbone network, combining depthwise separable convolution and channel shuffle mechanisms to significantly reduce the model’s parameter count and computational load during feature extraction. Second, a multi-branch feature fusion module, RepNCSPELAN4, was adopted, integrating group convolution and parameter-sharing strategies to balance global and local features, thereby enhancing multi-scale feature fusion efficiency. Then, a wavelet pooling (WaveletPool) layer was designed to reduce noise interference while preserving high-frequency texture details, improving the model’s ability to capture subtle bark texture features. Finally, the detection head structure Detect_Efficient was optimized using a dual-branch group convolution architecture to improve computational efficiency. Additionally, based on a self-built dataset of 70 tree species with 6 681 bark images, ablation studies and comparative experiments were conducted to thoroughly evaluate the performance of improved model.

Result The model achieved a detection precision of 98.1%, recall of 98.4%, mean average precision (mAP50) of 0.993, mean average precision across different IoU thresholds (mAP50-95) of 0.750, and an F₁ score (harmonic mean of precision and recall) of 0.982. Compared with the YOLOv11 model, the parameter count and computational load were reduced by 46.92% and 51.5%, respectively, significantly lowering the model’s spatial and computational complexity. It maintained stable identification performance under varying lighting conditions, demonstrating strong illumination robustness.

Conclusion The proposed YOLOv11-SWER model, through lightweight design and multi-scale feature optimization, achieves high detection accuracy while reducing parameters by nearly half, striking a good balance between high accuracy and efficiency. This method holds promise for applications in intelligent forestry monitoring and urban forestry resource management.