Abstract:
Abstract: An accurate and rapid extraction of fruit tree canopy has been one of the most important steps to determine the precise application of pesticides in modern agriculture. However, the current detection cannot fully meet the harsh requirements in the complex environment of an orchard in recent years. In this study, a novel extraction system was proposed to identify the fruit tree canopy in an orchard using millimeter-wave radar. Firstly, the region of interest (ROI) was set to remove a large number of environmental point clouds, and then the random sampling was used to fit and segment the ground point clouds. Secondly, a density-based clustering was selected to integrate the resolution of millimeter-wave radar and the distance parameter of the point cloud. An adaptive density clustering was then established to deal with the different densities of the point cloud at different distances when using the millimeter-wave radar. As such, the neighborhood radius was adaptively adjusted in different directions, according to the distance from the point cloud to the radar. The point cloud density was then used to identify the fruit tree canopy, further improving the recognition performance of the point cloud. Finally, an Alpha-shape was selected to reconstruct the three-dimensional surface of the fruit tree. The optimal parameters of the three-dimensional reconstruction were achieved to evaluate the volume from Alpha values and the geometric measurement. The random sampling was also used to extract the structural parameters. The parameters of the mathematical model were estimated from a data set with a large number of outliers using multiple iterations. A curve model was then obtained closer to the actual situation. In addition, a field test was conducted to verify the feasibility of the model in the Loquat Plant Resource Garden (113.367 789°E, 23.164 129°N) of South China Agricultural University on April 2th, 2021. 40 yellow-bark fruit trees were also scanned in the test. The results showed that a higher performance was achieved to effectively identify and extract the point clouds of a single tree canopy, where the F1 score was 93.7% for the recognition accuracy of fruit trees. Furthermore, the average relative errors of the plant height and crown width in the fruit trees were 8.7% and 8.1%, respectively, while the coefficients of determination were 0.84 and 0.92, respectively, and the root mean square errors were 16.39 and 7.82, respectively, compared with the manual measurement. In addition, the average volume of fruit trees was 5.6 m? using Alpha-shape, increasing by 59.4% in the accuracy of volume, compared with the traditional. Nevertheless, two recommendations can be addressed during this time: 1) To identify the fruit trees with the overlapping canopies using a relatively higher resolution of millimeter-wave radar; 2) To quantitatively extract the volume of the fruit tree in the harsh sense. Anyway, the millimeter-wave radar can be widely expected to accurately extract the canopy information in an orchard. The finding can provide a new promising technology to extract the canopy information for the data collection and automatic operation in modern agriculture.