| Lamiophlomis rotata(Benth.ex Hook.f.)Kudo is a perennial herb of the Lamiaceae family,mainly distributed in Qinghai,Tibet,Gansu,western Sichuan and northwestern Yunnan in China.It grows in the alpine degraded grassland at an altitude of 2700-4500m on the Qinghai-Tibet Plateau and is a first-class endangered Tibetan medicine.this study combines machine vision recognition with UAV technology,and uses artificial neural network algorithm to identify the Lamiophlomis of the community,improve the timeliness and accuracy of wild medicinal resources investigation.It provides a scientific basis for the dynamic management and sustainable utilization of Lamiophlomis resources at different spatial scales.Objective:1.By establish a method to use drones and deep learning Mask R-CNN model to quickly calculate the number of plants,leaf area and yield of Lamiophlomis.2.Through UAV remote sensing positioning,the spatial distribution pattern of the Lamiophlomis population was analyzed.Methods:1.Using the Lamiophlomis images obtained by UAV as the data set,compared the orthophoto maps assembled by Pix4Dcapture and Agisoft Metashape software,and all kinds of appearance and color of were compared.2.Labelme tool was used to label the Lamiophlomis images,and then used the Mask R-CNN model to train the Lamiophlomis images.Relevant parameters were adjusted according to the results,and compared the Lamiophlomis recognition rate of different degradation levels of grassland and drone flight height,and extracted the corresponding target feature recognitioned Lamiophlomis according to the three stages of images feature extraction-regional proposal network(RPN)-full convolutional network(FCN).Finally,in accordance with the process of‘cropping-recognition-splicing’,the whole remote sensing area was recognized by artificial intelligence.3.Envi5.3 software was used to count the information of in orthophoto map,including the number of plants and leaf area.The yield of was calculated according to the relationship between aboveground dry weight and leaf area.4.The model was used to identify the Lamiophlomis,and calculate its coverage and density,the point pattern method and Ripley’s L function were used to analyze the point pattern and population structure of Lamiophlomis in different degraded grasslands.Results:1.The results showed that Metashape software was slightly better than pix4d software.The survey area of the Ruoke River pasture sample in Aba County was9940m~2,the sample plot area of Qiongyue Mountain in Jiuzhi County,Qinghai Province was 2960m~2,and the plot area of Nianbaoyuze in Jiuzhi County,Qinghai Province was 3770m~2。2.The recognition rate of Lamiophlomis at different flight altitudes was compared by Mask R-CNN model,it was found that the average recognition accuracy(m AP)of was 0.8646 at 10m flight altitude,and the m AP of 15m was 0.8248.By compared the independent recognition rate under different grades of degraded grassland,the m AP of moderately degraded grassland was 0.8389,and that of severely degraded grassland was 0.7113.In all data sets,when IOU was greater than 0.5,the m AP in the training set was 0.8679,and the m AP in the verification set was 0.8022.The recognition in the orthophoto maps showed that the Lamiophlomis recognition rate was greater than 90%.3.Field investigation found that the linear relationship between dry weight and leaf area was y=0.53529x-0.00154.Envi software segmentation and statistical results showed that in the flight image of the distribution area of Lamiophlomis.It was concluded that there were 22631 plants of Lamiophlomis in the moderately degraded grassland in Ruoke River pasture sample,with a leaf area of 207.1m~2 and a yield of110.87kg.The number of Lamiophlomis in Qiongyue mountain sample was 14,818,the leaf area was 107.8m~2,and the yield was 57.7kg.The number of Lamiophlomis in the Nianbaoyuze sample was 28,262,the leaf area was 183.0m~2,and the yield was 98.0kg.4.The results of the spatial distribution pattern of the Lamiophlomis population showed that in the moderately degraded grassland,the cover degree was 5.39%,and the density was 10.1 plants/m~2.The cover degree of the severely degraded grassland was 3.4%,and the density was 8.4 plants/m~2.The point pattern analysis results showed that in the sample plots of moderately and severely degraded grasslands and edge plots,it was found that the Lamiophlomis populations in most of the plots were clustered,and the maximum scale of clustering distribution was about 6m-7m,a small part of the plots were clustered distribution-random distribution.Conclusion:This research uses UAV remote sensing and artificial intelligence technology to establish a method for rapid plant identification and calculation of Lamiophlomis plant number,leaf area,yield,and population spatial distribution pattern analysis in the remote sensing area.It has realized the rapid and accurate estimation of Lamiophlomis wild resource reserves and the study of population spatial distribution pattern,effectively improved the efficiency and accuracy of resource investigation,provided a new technology and method for medicinal plant resource investigation. |
PDF Full Text Request