Banana Individual Segmentation and Phenotypic Parameter Measurements Using Deep Learning and Terrestrial LiDAR

Abstract

Banana phenotypic parameters are one of the important elements in the study of banana growth and development. Through the measurement of banana phenotypic parameters, we can obtain information about the growth status, nutritional status and quality indexes of banana plants, and pseudo-stem parameters are significant indicators in banana phenotypic parameters. This research proposes a two-stage approach combining morphological features and deep learning point cloud segmentation to extract banana pseudo-stem parameters. Specifically, in the first step, seed points are extracted using the DBSCAN clustering algorithm, and banana individual plant segmentation is accomplished using the region growing algorithm based on seed points. Its precision, recall and F1-score were 97.73%, 97.36% and 97.54%, respectively. This indicates that the DBSCAN clustering algorithm and the seed point based region growing algorithm can effectively realize the plant count of banana plants and initially realize the individual plant segmentation of banana. The second step is to use PointNet++, PointNet, and DGCNN for pseudo-stem and canopy segmentation of individual banana plants. All three models perform well in segmentation, with PointNet++ performing the best. Its precision, recall, F1-score, Matthews correlation coefficient and Dice coefficient reached 0.9956, 0.9709, 0.9831, 0.9670 and 0.9831. This shows that deep learning has a better applicability in segmenting banana plants. In the results of segmentation, we measure the banana pseudo-stem circumference and pseudo-stem height. The correlation between the extracted pseudo-stem height and pseudo-stem circumference compared to the measured values was 96.70% and 82.32%, respectively. The above two-stage method of extracting banana pseudo-stem parameters overcomes the difficulties of point cloud individual plant segmentation associated with intensive banana cultivation. It makes the management of individual banana plants possible and provides accurate phenotypic parameter information for banana plantation management. It lays the foundation for further assessment of banana growth and nutritional status.

Keywords

• Terrestrial LiDAR
• individual segmentation
• phenotypic parameter
• deep learning