Optimal phase and method for identifying pear trees in Guanzhong area via remote sensing

Abstract:【Objective】In the present study, 22 GF6-WFV images were used to explore the optimal phase and method for identification of pear trees (including Dangshansu pear trees and Zaosu pear trees) by remote sensing in Guanzhong area, in order to provide a method for monitoring of pear trees using remote sensing.【Methods】Firstly, each image was preprocessed (including image space clipping, image radiation calibration, image atmospheric radiation correction, image geometry correction, image mean filtering, etc). Then, the identification efficiency of eight methods (including red edge parameter method, spectral distance method, image enhancement processing and analysis method, image difference and ratio method, reflection spectrum and bands difference method, spectral indices and their change analysis method, and optimal combination of identification methods) were tested based on the region of interest (ROI) data of sample plots of 15 crops. The better methods and their corresponding application phases were optimized. Finally, the identification accuracy and solidity of these methods and their optimal combination were verified by using global images.【Results】(1) The best identification phase to identify pear trees was the full flowering stage, and the identification accuracy at fruit ripening stage and other phases was not ideal. (2) The RGB component threshold method had strong identification effect on pear trees in full blooming period (the RGB component threshold method refers to that the images of each phase were processed by false color synthesis; then, the obtained false color synthesis images were processed by the method of grayscale stretch, and the stretched results were stored as 24 bites RGB images; Finally, the differences of RGB component vales between pear trees sample pixels and those of non pear crops were compared and analyzed). The RGB component values of sample pixels of pear trees and the other crops were of great differences when the extreme values of R, G and B component of pear sample pixels were used as thresholds (i.e. R component≥255, 0≤G component≤ 161, 143≤B component≤255). It was found that there were relatively more misclassification pixels in plum, cherry, kiwi and winter rape (the misclassification rates were 36.03%, 4.92%, 5.38% and 74.92%, respectively), and the misclassification pixels were mainly concentrated in a few plots and were fewer in the other 10 non pear crops (the misclassification rates＜4.92%) and scattered. The overall accuracy (calculated according to the two categories of pear trees and non pear crops) was as high as 94.28%. (3) The R710 threshold method (R710 refers to the reflectance of the band with the central wavelength of 710 nm) also had a strong identification effect on pear trees in full blooming period. Its identification accuracy was higher than the common vegetation indices (such as R710-R425, MSRred- edge=(R750-R425)/(R710 + R425), SRred- edge= R830/R710, CLred-edge=R830/R710-1, IRECI=(R830-R660)/(R710/R750), NDVI=(R830-R660)/(R830+R660), NDVIred- edge=(R750-R710)/(R750+R710), mNDVIred- edge=(R750-R710)/(R750+R710-2×R425), etc). The R710 of pear sample pixels was higher than that of most non pear crop sample pixels. When the extreme values of R710 of pear tree sample pixels were used as the threshold (i.e. 0.234 0≤R710≤0.264 1), it was found that there were relatively more misclassification pixels of plum, peach, pomegranate, persimmon, apple and grape (the misclassification rates were 21.45%, 74.97%, 25.22%, 39.84%, 9.64% and 100%, respectively), and there were fewer misclassification pixels of the other eight non pear crops (the misclassification rates were≤0.15%) and the overall accuracy was 79.78%. (4) It was impossible to distinguish pear trees from plum trees only by using the image of pear trees in full blooming period, but using the threshold value of the difference of red edge 1 band (R710-apr-R710-mar) in the images of pear flowering and plum flowering, pear and plum trees could be accurately distinguished. The value range and mean value of R710-aprR710- mar corresponding to plum sample pixels were 0.058 1-0.087 9 and 0.071 3, respectively; the value range and mean value of R710- apr-R710- mar corresponding to pear sample pixels were 0.082 9-0.113 9 and 0.095 7, respectively. When taking the minimum value of R710- apr-R710- mar of pear sample pixels as the threshold (i.e. R710- apr-R710- mar≥0.082 9), it was found that although the overall accuracy of this method was not high it could reduce the misclassification rate of plum sample pixels to 6.81% (The misclassification rates of plum sample pixels corresponding to other methods in this paper was≥21.45% ). (5) There was certain complementary effect among the three threshold methods including RGB component, R710 and R710-apr-R710-mar, and the decision tree constructed by their combination had the best identification effect on pear trees, with a correct rate of pear trees of 92.91%, that of non-pear crops up to 97.53% and the overall accuracy of 97.19%.【Conclusion】The decision tree based on the three threshold methods (including RGB component, R710 and R710-apr-R710-mar) can identify pear trees by using images at blooming period with a high accuracy in the study area.