Automated extraction of tree and plot-based parameters in citrus orchards from aerial images

Abstract

A plot-based approach is proposed to detect fruit trees from high spatial resolution aerial images and extract tree and plot-based parameters, such as fraction of tree cover, number of trees, and planting patterns. Each plot image, defined by the boundaries or polygons obtained from a cadastral database, is analyzed independently. The methodology is based on image processing methods: an unsupervised classification with the k-means algorithm is applied, followed by the automatic identification of the classes representing trees. Once extracted, each tree is individualized using a morphological process applied on the binary image of the trees. A set of parameters is calculated at tree and plot levels that produces a comprehensive description of the spectral and morphological aspects of the trees, as well as their spatial distribution in each plot. The methodology was tested on 0.5 m/pixel spatial resolution aerial images of 300 citrus orchard plots which included the three citrus tree typologies found in the Valencia region (Spain). The accuracy of the fruit tree extraction and the parameters calculated was evaluated by comparison with reference data obtained by manual delineation of the images. The automatically extracted fraction of tree cover was significantly related to the reference tree cover area (R2 = 0.96). In the case of the number of detected trees, the R2 values were always higher than 0.90 for the three typologies. Tree location was estimated with an average positional error of 40 cm. The error obtained in the characterization of the planting pattern was less than 50 cm. The proposed methodology may be applied to large agricultural databases, and the derived information combined with precision agricultural techniques could improve the efficiency of various irrigation and agricultural management tasks such as handling per-plot water requirements and distribution.