基于实例分割的高郁闭度林分单木树冠无人机遥感提取

Tree Crown Extraction of UAV Remote Sensing High Canopy Density Stand Based on Instance Segmentation

[目的]利用遥感影像获取高郁闭度林分树冠信息。[方法]试验了一种基于实例分割模型的无人机遥感影像单木树冠提取方法,选用7种残差网络用于模型的特征提取,逐一对不同郁闭度杉木纯林进行单木树冠提取。[结果]表明,7个实例分割模型对低郁闭度林分树冠分割的边界框AP值和掩膜AP平均值分别为55.89%、57.29%,林分东西冠幅、南北冠幅和树冠面积参数提取均方根误差平均值分别为0.161、0.179和0.341,平均预测决定系数R~2分别为0.912、0.918和0.957;对高郁闭度林分树冠分割的边界框AP值和掩膜AP平均值分别为46.00%、44.45%,单木东西冠幅、南北冠幅和树冠面积参数提取均方根误差平均值分别为0.479、0.497和1.256,平均预测R~2分别为0.806、0.762和0.936。[结论]各参数提取精度均优于传统调查精度,该方法能自动化、快速化、精准化获取树冠信息。

[Objective]To obtain canopy information in high canopy density forest by remote sensing images.[Method]A single tree crown extraction method of UAV remote sensing image based on instance segmentation model was tested.Seven residual networks were selected for feature extraction of the model,and the single tree crowns of pure Chinese fir forests with different canopy density were extracted one by one.[Result]The results showed that the average boundary AP value and mask AP value of seven instance segmentation models for canopy segmentation of low canopy density forest were 55.89%and57.29%,respectively.The average RMSE of east-west crown width,north-south crown width and crown area parameters was 0.161,0.179 and 0.341,respectively.The R2 was 0.912,0.918 and 0.957,respectively.The average boundary AP value of canopy segmentation and the average AP value of canopy cover of high canopy density forest were 46.00%and 44.45%,respectively.The average RMSE of east-west crown width,north-south crown width and crown area parameters was 0.479,0.497 and 1.256,respectively.The average predicted R2 was 0.806,0.762 and 0.936,respectively.[Conclusion]The extraction accuracy of each parameter is higher than the traditional survey accuracy,and this method can obtain crown information automatically,rapidly and accurately.