| Plant phenotypic traits are closely related to plant growth state and yield,which can provide comprehensive basis and guarantee for plant growth monitoring and research.Compared with traditional manual measurement methods,UAV remote sensing can obtain data more efficiently.However,in the face of complex and diverse surface habitats,how to use the characteristics of relevant data sources? To identify and accurately describe crop phenotypic traits from a variety of surface habitats has become a problem to be solved.In this study,Pitaya plants,a typical cash crop in the complex growing environment of Guanling-Zhenfeng Huajiang Demonstration area,were taken as the research object,and RGB images and photographic measurement point cloud data were collected by UAV as data sources.Combined with the field measurement data of surface habitat,the heterogeneous surface landclasses were divided according to the complexity of pitaya’s habitat,in order to summarize the characteristics of relevant data and phenotypic traits of pitaya plants in different habitats.In order to draw relevant conclusions according to the characteristics of different surface habitats.The features of different data sources were used to calculate the vegetation index and segment the intersection threshold of RGB images.The point cloud data was interpolated for CHM segmentation.At the same time,the multi-source feature fusion data was segmented by object-oriented rule method.This paper attempts to compare the optimal data sources and methods for plant extraction and description of different habitat types in heterogeneous land surface from the aspects of single data and multi-source data characteristics.The main research contents and results are as follows:(1)The heterogeneous surface of plants was divided into plant-bare soil,plant-weed,plant-ladder ridge-bare soil and plant-ladder ridge-weed,according to different habitat types,and the surface sample of typical habitats was selected for data analysis and field measurement.In combination with the ROI training sample area data,ground object pop profile rendering and reflection value statistical analysis were carried out on the RGB image of the sample area.The results confirmed that in the surface of the four habitats,the reflection spectra of plants and weeds were in the same trend,the overall reflection values were close,and the reflection values of ground objects in the vegetation area and non-vegetation area were significantly different.(2)The four typical vegetation indices of VDVI,NGBDI,NGRDI and RGBVI were calculated for RGB images of the four habitats,and the statistical values of ROI data were used for histogram rendering.The threshold intersection method was used for plant identification and extraction combined with the statistical values.The results were quantitatively calculated and analyzed,and the field measurements were compared at the same time.The results showed that the extraction results of VDVI index were more consistent with the real plant contour and had good segmentation effect.In the two habitats containing weeds,plant-weed and plant-step-ridge and weed,there was a large area of weed connection,and it was impossible to separate plants from weeds.The overall accuracy F measure of single plant matching in plant-bare soil and plant-ladder ridge-bare soil environment without weeds was 76.46% and61.18%,respectively.The interference factors were mainly caused by branch connection between plants and few weeds.The difference between the extracted crown width and contour area of a single plant and the measured value was small,and the extracted value was slightly larger than the measured value,which was mainly affected by the transition zone of RGB raster image.The Pearson correlation coefficients of crown width and contour area with measured values reached 0.884 and 0.887,respectively,reaching a highly correlated state,and the overall description was accurate.(3)The point cloud data are pre-processed,such as de-noising and ground point separation,and then the digital surface model and digital elevation model are obtained by interpolation.The difference between the two can be used to obtain the canopy height model CHM of surface objects,and the plant height information is used for segmentation,extraction and analysis.The results showed that in the plant-weed habitat,the overall accuracy of plant number identification was up to 80.63%,and the recall P value in the plant-bare soil habitat was 45.25%,which was mainly caused by the short growth years of plants in the plant-bare soil habitat and the low density of the broken branches.In the plant-weed habitat,the dense extraction effect of point cloud data was better because of the high density of vegetation.The misjudgment rates in plant-ladder ridge-bare soil and plant-ladder ridge-weed habitats were both high,reaching 42.86% and 36.42%,respectively,which led to an increase in the number of overall segmentation,and then reduced the overall accuracy F measure to 62.75% and 67.63%.The main reason is that the two habitats contain the "ladder ridge" factor which is similar to the height of the plants,leading to the formation of false segmentation.(4)Because the single data RGB image and point cloud data were affected by weed and ladder ridge factors,the results showed certain errors in the number of plants and the description results of phenotypic traits of individual plants.So the advantages of the two were integrated and the object-oriented rules of multiple data sources were used to carry out multi-feature fusion,so as to remove the interference factors to the maximum extent.At the same time,the result features of single data extraction were compared.The results showed that the feature extraction method of multi-source data significantly improved the number identification of plants in two habitats: plant-ladder ridge-bare soil and plant-ladder ridge-weed.When the influence of ladder ridge factor was excluded,the misjudgment rate decreased significantly,and the F measure increased to 93.48% and 91.89%,and the extraction accuracy increased significantly.In the plant-bare soil environment,the results of the fused multi-source feature data were similar to the extraction results of the point cloud data,and the matching accuracy was not high because of the low vegetation density.In the three environments,plant-weed,plant-ladder ridge-bare soil and plant-ladder ridge-weed,there was little difference between the description parameters and the measured values,and the Pearson correlation coefficient was high,and the overall results could better describe the surface characteristics of plants.For the surface habitat without vegetation,it is obvious that the description of plant micro-scale phenotypic traits based on RGB images can better represent the real plant contour characteristics,and for the environment containing ladder and ridge elements.The feature method of integrating multi-source data is more representative for the identification of the number of macro-scale plants. |
