Deep Learning-based Object Detection for Crop Monitoring in Soybean Fields

In this paper, a soybean flower/seedpod detection system is built for collecting growing state information by introducing convolutional neural networks, aiming that observed plant states (e.g., #flowers and #seedpods) are used to predict the crop yields of soybeans by combining the environment information in future. To predict the crop yields (i.e., quantity of seedpods) precisely, it is considered important to know how the number of flowers are translated over time and how such flower transients can affect the final yields of soybeans. However, there has not existed a way to measure the number of flowers in real environments. For this purpose, We propose a deep learning approach to automatically detect flower and seedpod regions from images which are taken in real soybean fields without environmental control. Various object detection methods are compared, including RetinaNet, Faster R-CNN, and Cascade R-CNN. Ablation studies are provided to analyze how these methods perform on both flower and seedpod across different parameters. In our experimental results, Cascade R-CNN gives the best average precision (AP) of 89.6, while RetinaNet and Faster R-CNN give AP of 83.3 and 88.7, respectively. Cascade RCNN also attains the highest accuracy in detecting small objects, which are not easily detected by other models. With accurate detection, the system is expected to contribute to constructing high-performance measurement for soybean flowers and seedpods, which ultimately leads to better pipeline in evaluating plant status.