Hierarchical graph representation for unsupervised crop row detection in images

Abstract

Crop row detection is an important aspect of smart farming. An accurate method of crop row detection ensures better navigation of the robot in the field as well as accurate weed control between rows, and crop monitoring. Recent deep learning approaches have emerged as the indispensable methods for most computer vision tasks. Therefore, crop row detection using neural networks is currently the most frequently studied approach. However, these methods require a large amount of labeled data, which is not suitable for many situations such as agriculture, where labeling data is tedious and expensive. Unsupervised techniques such as graph-based represent a promising alternative to tackle such a problem. Indeed, the crop rows represent relationships between plants (relative position, co-occurrence...), that could be integrated as structured information with an unsupervised approach. However, little attention has been paid to graph-based techniques for crop row structures representation. In this paper we propose a new method based on hierarchical approach and unsupervised graph representation for crop row detection. The idea is to transform the field into a structured data where each plant can be linked to its neighbors according to their spatial relationships. Then, with a hierarchical clustering and a criterion of exploring the nodes of the graph we extract subgraphs that represent the aligned structures. We show that the graph matching technique can discard inconsistent structures such as weed aggregations. The results obtained show the validity of the proposed method, with higher performances.