Graph-Based Deep Learning for Component Segmentation of Maize Plants

TL;DR

This paper introduces a novel graph neural network architecture utilizing GNNs, PCA, and attention mechanisms to improve the segmentation of maize plant components from LiDAR 3D point cloud data, achieving over 80% IoU accuracy.

Contribution

The work presents a new graph-based deep learning model that effectively segments plant components in 3D LiDAR data, outperforming existing methods.

Findings

Achieved over 80% IoU in component segmentation

Enhanced accuracy over traditional point cloud models

Demonstrated effectiveness of GNNs with PCA and attention mechanisms

Abstract

In precision agriculture, one of the most important tasks when exploring crop production is identifying individual plant components. There are several attempts to accomplish this task by the use of traditional 2D imaging, 3D reconstructions, and Convolutional Neural Networks (CNN). However, they have several drawbacks when processing 3D data and identifying individual plant components. Therefore, in this work, we propose a novel Deep Learning architecture to detect components of individual plants on Light Detection and Ranging (LiDAR) 3D Point Cloud (PC) data sets. This architecture is based on the concept of Graph Neural Networks (GNN), and feature enhancing with Principal Component Analysis (PCA). For this, each point is taken as a vertex and by the use of a K-Nearest Neighbors (KNN) layer, the edges are established, thus representing the 3D PC data set. Subsequently, Edge-Conv layers are used to further increase the features of each point. Finally, Graph Attention Networks (GAT) are applied to classify visible phenotypic components of the plant, such as the leaf, stem, and soil. This study demonstrates that our graph-based deep learning approach enhances segmentation accuracy for identifying individual plant components, achieving percentages above 80% in the IoU average, thus outperforming other existing models based on point clouds.