Evaluating current techniques for enhanced precision in automated soybean seed counting

Agriculture is one of the oldest activities in the world and has always been at the forefront of technological innovation. With mechanical equipment, modified seeds, and digital devices, every aspect of farming, from planting to harvesting, is gradually being improved. These benefits have also translated into better appreciation for crop yields such as soybeans.

Deep learning-based yield estimation models use methods such as regression, traditional bounding boxes, or density maps to make seed counting easier. Compared with manual counting, these methods are undoubtedly simpler, more accurate, and easy to implement.

“P2PNet” is one of the automated counting methods recently proposed to simplify point counting for soybean seeds. However, this method showed low performance for direct seed counting. Perturbation from background objects, large predictions, use of high-level features, and uncalculated object scale were identified as some of the drawbacks of this model.

To address the challenges associated with this paradigm, researchers from Japan have developed a new paradigm that adds to the list of agricultural technological innovations. Accurately counts the number of soybean seeds from field photos of soybean plants, eliminating the labor-intensive process of seed counting.

The study was led by Associate Professor Wei Ju of the University of Tokyo and is published online in plant phenomics.

“Soybeans are an important source of protein for animals and humans. Therefore, achieving high crop yields is a common standard and goal in most breeding programmes,” explains Professor Guo.

Seed count is of particular interest to growers as it can be used to ascertain both the yield of a plant and its propagation potential. Traditional image-based automated seed counting methods track seeds in images by placing them in bounding boxes. However, in actual field conditions, the presence of complex backgrounds, overlapping pods, and variable light conditions can cause overlapping of bounding boxes, resulting in inaccuracies in seed number and placement.

To address these challenges, the team upgraded P2PNet to the new and improved “P2PNet-Soy” model. Counts objects by selecting them as small dots in the image. To obtain data to train the model to identify soybean seeds, the researchers took 374 images of soybean plants grown in a field.

They photographed two sides (front and back) of the plant to capture the maximum amount of seeds in the plant. Then, skilled technicians from the University of Tokyo’s Phenomenology Laboratory carefully marked the seeds in each soybean with dots. Ensure that only seeds belonging to the target plant are annotated and exclude seeds from neighboring and background plants. The researchers then selected 181 training images and used another 193 images taken from the contralateral side to evaluate the model.

The researchers adopted several strategies to improve the performance of the model. First, high and low level features were captured from field images. High-level features generally consider the context of objects in images, while low-level features are more useful for recognizing details and smaller objects.

An unchanging feature extraction method known as atrial torsion was then used to enable the model to detect seeds of different sizes. In addition, spatial attention and channel mechanisms have been applied to better distinguish seeds from background. The research team refined the model’s predictions by applying a post-processing technique called kd tree, which is an unsupervised clustering algorithm that closely identifies the centers of predicted seed locations, which enhances the accuracy of the final prediction.

These improvements resulted in an accurate seed counting and localization model that can detect and count seeds from simple images of soybean plants taken in the field. “The upgraded P2PNet-Soy method for more efficient soybean seed counting and localization has a much higher accuracy compared not only to the original P2PNet but also other soybean counting method,” says Professor Gu.

Although these improvements improve the accuracy of seed prediction, the model has some limitations that need to be fixed. Because the model is trained on images taken from both sides of the same plant, it can overestimate the number of seeds on the plant. In addition, the model cannot detect seeds that have been accidentally lost in the image.

However, the development of such advanced technologies is a promising step towards a more efficient agricultural industry.