基于改进YOLOv8的无人机遥感影像树种识别方法

Tree Species Recognition Method from UAV Remote Sensing Images Based on Improved YOLOv8

为利用周围环境信息识别不同树种,建立松树(Pinus spp.)、杉木(Cunninghamia lanceolata)、桉树(Eucalyptus spp.)和其他阔叶树的高分辨率无人机影像数据集,验证YOLOv8-LSK算法的识别效果;通过公共数据集,验证YOLOv8-LSK算法的泛化能力;为验证YOLOv8-LSK算法的精确率,将YOLOv8-LSK算法与5种算法进行对比;通过消融试验,验证YOLOv8-LSK算法的有效性。为验证空间注意力效果,采用YOLOv8算法作为基线,将LSK模块与不同轻量级模块进行比较。结果表明,与R3Det、CFA、AOPG和RVSA算法相比,YOLOv8-LSK算法mAP最高(81.23%),泛化能力较高。与TridentNet、RT-DETR、ReDet、Faster-RCNN和RTMDet算法相比,YOLOv8-LSK算法mAP最高(78.61%),精确率较高。消融试验结果表明,与YOLOv8、YOLOv7和YOLOv6算法相比,YOLOv8-LSK算法mAP显著提升。与CBAM、SKNet和ConvNext模块相比,YOLOv8-LSK算法mAP最高(78.61%)。YOLOv8-LSK算法识别的树种图斑边界均较明显,感受野更大。

In order to identify different tree species by combining information of surroundings,high-resolution Unmanned Aerial Vehicle(UAV)image datasets of Pinus spp.,Cunninghamia lanceolata,Eucalyptus spp.and other broad-leaved trees were established to verify recognition effect of YOLOv8-LSK algorithm.Generalization ability of YOLOv8-LSK algorithm was verified by public dataset.To verify accuracy rate of YOLOv8-LSK algorithm,YOLOv8-LSK algorithm was compared with 5 algorithms.Effectiveness of YOLOv8-LSK algorithm was verified by ablation experiment.To verify spatial attention effect,YOLOv8 algorithm was used as baseline,and LSK module was compared with different light weight modules.Results showed that compared with R3Det,CFA,AOPG and RVSA algorithms,YOLOv8-LSK algorithm had higher generalization ability,with the highest mAP of 81.23%.Compared with TridentNet,RT-DETR,ReDet,Faster-RCNN and RTMDet algorithms,YOLOv8-LSK algorithm had higher accuracy rate,with the highest mAP of 78.61%.Ablation experiment results showed that YOLOv8-LSK algorithm had significantly higher mAP compared with YOLOv8,YOLOv7 and YOLOv6 algorithms.Compared with CBAM,SKNet and ConvNext modules,YOLOv8-LSK algorithm had the highest mAP(78.61%).Boundaries of tree patches identified by YOLOv8-LSK algorithm were more obvious and receptive fields were larger.