Ridge-Structure-Based Crop Classification for Small UGV and Derived Detection Framework

Abstract

In crop detection, ridge structures provide crucial cues for classifying crops and weeds. However, it is difficult to obtain ridge structures for unmanned ground vehicles which can capture images only within a narrow field of view. This study proposes a lightweight algorithm that enables a model to implicitly infer the ridge structure from plant-to-plant spatial relationships and sizes. An object detector first detects each plant. The resulting bounding boxes are treated as pairwise features in the nodes. Metainformation indicating whether two nodes share the same ID is combined with their geometric relationships and encoded as edge features. A graph attention network addresses these relationships to infer and propagate ridge-aware regularities. By understanding the structure only from object relationships, the method compensates for the information lost to the limited field of view without any explicit edge structure input. In the experiments wherein we deliberately introduced a domain shift between the training/validation sets and test set, the proposed method increased the baseline mAP₅₀ from 30.6% to 44.4%. This amounts to an increase of up to 13.8 percentage points. In addition, the proposed method requires only approximately 10 ms/frame on a Jetson AGX Orin to classify plants. This method acquires ridge structures internally without relying on external sensors or hand-tuned thresholds. Thus, it displays potential for in-field agricultural applications such as autonomous weeding.

Overview of the proposed framework