Wheat Spikes Height Estimation Using Stereo Cameras

There is a positive correlation between wheat plant height and lodging, yield, and biomass. So, in precision agriculture, a high-throughput estimation of the wheat plant's height in terms of its spikes is essential. This study aims to develop a straightforward, cost-effective method for measuring the height of wheat plants using stereo cameras. To collect the required datasets, we conducted an experiment in which we collected RGB images along with their depth layer using two renowned stereo cameras, OAKD and D455. Then, we used a deep learning model called mask region-based convolutional neural networks to localize and distinguish the spikes in the collected images. In this study, we localized the wheat spikes using object detection (OD) and instance segmentation (IS) models. The advantage of the OD model over the IS model is that its bounding box annotation procedure in the data preparation phase is significantly faster than the IS model's polygon annotation. However, the disadvantage of OD is that there are many background pixels in each predicted bounding box, which reduces the performance of height estimation. To facilitate the annotation process of the collected datasets, we also developed a hybrid scale-invariant feature transform random forest-based active learning algorithm to transfer the annotations of one camera to the other. The results show that the OAKD camera performs better than the D455 camera for wheat plant height estimation due to its higher RGB quality and better matching of the mono camera images. Using the OAKD camera and IS model, the algorithm proposed in this study is able to estimate wheat height with mean absolute percentage error values of 0.75% and 0.67% at the spike and plot levels, respectively.