TROPICAL GEOGRAPHY ›› 2019, Vol. 39 ›› Issue (4): 502-511.doi: 10.13284/j.cnki.rddl.003146

Previous Articles     Next Articles

Recognition and Extraction Method of Single Dragon Fruit Plant in Plateau-Canyon Areas Based on UAV Remote Sensing

Zhu Menga,b, Zhou Zhongfaa,b,c, Zhao Xina,b, Huang Denghonga,b, Jiang Yia,b, Wu Yuea,b and Cui Liangd   

  1. (a. Karst Research Institute//Department Geography and Environment Sciences, Guizhou Normal University; b. State Key Laboratory Incubation Base for Karst Mountain Ecology Environment of Guizhou Province; c. State Engineering Technology Institute for Karst Desertification Control; d. Guizhou Beidou Space Information Technology Co. Ltd., Guiyang 550001, China)
  • Online:2019-07-10 Published:2019-07-10


Based on a small-sized Unmanned Aerial Vehicle (UAV) with a visible light lens for data acquisition, and with the flight height set to 80 m, the heading overlap is 75% and the side overlap is 70% in obtaining centimeter-level visible images. The five types of visible spectral-index based methods including Visible-Band-Difference-Vegetation-Index (VDVI), Excess-Green index (ExG), Excess-Green minus Excess-Red index (ExG-ExR), Normalized Green-Red Difference Index (NGRDI), and the Green Leaf Index (GLI) were applied to distinguish pitaya fruit of typical economic crops in karst plateau gorge areas. The results showed that VDVI, ExG, and ExG-ExR are effective in separating target features from the soil, gravel, ridge, and subtilis, while GLI and NGRDI completely intersect, and the overlap phenomenon is noticeable with the background pixels. In computing and comparing the suitability of Otsu and bimodal histogram threshold extraction methods, VDVI, ExG and ExG-ExR display noticeable bimodal effects, while the histogram bimodal effect of NGRDI and GLI is not prominent, and the calculation result is still a qualitative description. The maximum class variance algorithm can be used to quantitatively obtain the threshold of the target and the background. Comparing the recognition characteristics of the five indices, the number of pitaya fruit pixels were extracted by VDVI accounted for 13.94% of the whole study area, the recognition accuracy was 97.437%, the Kappa coefficient was 0.9607, the recognition of pitaya fruit was higher, and the resolution was also good. The accuracy of GLI and ExG-ExR is only 42.57% and 58.96%, respectively. Clearly, the two indices misclassify other features or background values as target features. The target pixel scale of NGRDI is only 6.61%, and most of the pitaya fruit area is missed. ExG has evident recognition and extraction ability compared to ExG-ExR, NGRDI, and GLI, but there is still a certain missed phenomenon compared to VDVI. Combined with visual space modeling tool Model Builder, proposed a method of slicing the plant clusters by the average area of a single plant. Pitaya fruit of typical economic crops was used as the research objects for conducting plant recognition and information extraction methodology research. In conclusion: 1) Compared to EXG, NGRDI, ExG-ExR, and GLI vegetation indices, VDVI identified 161 species of dragon fruit, which reached 50.31% of the total number of plants in the study area and is the best segmentation method for dragon fruit recognition; 2) Compared to the five indexes, it could well segment the target and background objects when the VDVI segmentation threshold is 0.037; 3) Through the verification of the actual number of plants, and the number of extracted plants measured in the human–machine interaction base, the overall precision of the single plant extraction of pitaya fruit was 91.7%. The results confirmed that it is feasible to use visible light images as the data source, the VDVI as the identification method, and for plant cluster segmentation to be applied in the identification of dragon fruit in karst plateau gorge areas.

Key words: UAV remote sensing, crop identification, threshold extraction, single plant extraction, color index, Kast Plateau Canyon area