Study On The Extraction Method Of Dead Trees From Shelter Forests Based On UAV Remote Sensing

As an important barrier of the agricultural ecosystem,the monitoring of the damage and health status of the shelter forest is especially important in the management of the northern forest network in China.Dead trees are mostly caused by diseases,water shortage and unreasonable planting structure.Therefore,it is crucial to find an accurate and efficient method to extract dead trees from shelter forests.At present,remote sensing technology is mainly used for large scale forest network monitoring,but satellite remote sensing,which usually has low resolution and poor effectiveness,is gradually replaced by UAV highresolution remote sensing images with flexible flight time and large acquisition range.High-resolution remote sensing images have become a reliable data source for forest network monitoring.In this thesis,the UAV remote sensing platform is used to obtain the shelter forest images in the study area,the dead tree extraction of shelter forest based on multi-source remote sensing data fusion,the dead tree extraction of shelter forest based on small target lightweight target detection,and the dead tree extraction model of shelter forest based on multi-spectral waveband selection and instance segmentation are constructed respectively,and the research work carried out in this thesis are as follows:(1)Dead Tree Extraction from Shelter forests with Multi-source Remote Sensing Data Fusion.To address the problem of insufficient information of single remote sensing data,we propose a method of extracting dead trees from shelter forests by fusing multi-source remote sensing data,making full use of the3 D point cloud reconstructed from visible images and the spectral information of multi-spectral images.The ground point cloud is filtered and interpolated with Radial Basis Function Neural Network(RBFNN)to obtain Digital Elevation Model(DEM),and the original point cloud is interpolated to generate Digital Surface Model(DSM),Canopy Height Model(CHM)is obtained by subtracting the two.Then,the objectoriented method based on Support Vector Machine(SVM)is used to extract the canopy from CHM +multispectral raw band data,and compared with the canopy extraction method based on local maximum +watershed method for CHM.The result shows that the proposed method can effectively reduce the over segmentation and under segmentation problems,and the F1 value is increased by 0.12-0.17.(2)Small target-based lightweight target detection for shelter forest dead tree extraction.Aiming at the problem that remote sensing intelligent algorithms are usually limited by computational resources and need to balance the performance and efficiency of the algorithm,while the shelter forest as a small target cannot be well identified,an improved YOLOv5 network model applicable to small target detection is proposed.Specifically,introducing the dense connection idea and Soft Pool pooling method in the feature extraction model can effectively retain the detail information of small targets and reduce the missed score.The traditional convolution is modified into depth-wise separable convolution,which can reduce the parameter computation,while the feasibility of the method is verified by ablation experiments.The result shows that the improved YOLOv5 model can better balance the algorithm performance and detection efficiency when compared with advanced target detection methods,with an AP of 89.11%,a model parameter count of only7.6 MB,and a training time of 65 s per epoch.(3)Shelter forest dead tree extraction model based on multispectral waveband selection and instance segmentation.For UAV remote sensing images,it is difficult to obtain samples with high confidence,and there is no corresponding shelter forest dead tree dataset at the spectral level,and the means of shelter forest dead tree extraction with few samples is not yet comprehensive.The optimal band combination of multispectral data(without radiation correction)is selected based on the best index factor,and the dataset is labeled based on band 5,band 3,and band 1.The top 10 band combinations are exported to make the COCO dataset,and the dead tree extraction of the shelter forest is completed based on the SOLO model,and the SOLO model based on ResNet101+FPN backbone network achieves better extraction accuracy The AP was 61.3%(for the dataset containing only dead tree labels)and 63.8%(for the dataset containing both dead tree and healthy tree labels),and it was also proved that adding healthy tree samples could effectively improve the accuracy of dead tree extraction.The result shows that the difference of extraction accuracy did not exceed 2% with or without radiation correction,so whether or not radiation correction would not significantly affect the extraction accuracy of dead trees in shelter forests.