The crop phenotyping team at Huazhong Agricultural University (HZAU) has published a research paper titled "EU-GAN: A root inpainting network for improving 2D soil-cultivated root phenotyping" in the journal Artificial Intelligence in Agriculture.

The team has developed a GAN-based 2D root image restoration model that significantly enhances the accuracy of root phenotyping analysis.

The commonly used rhizobox method is a low-cost, nondestructive soil-cultivated root phenotypic detection technique. However, some roots are hidden in the soil, leading to data bias, and annotating the restoration dataset presents challenges. Current restoration methods involve using a large number of non-real images for restoration, resulting in limited effectiveness in repairing roots with long distances and edge deficiencies.

Root inpainting network architecture. [Photo/news.hzau.edu.cn]

The HZAU team has developed a fully open-source hybrid root inpainting dataset (HRID). By annotating a large number of soil-cultivated root images and using a hydroponic root box root addition algorithm to generate missing roots, the researchers created a dataset of realistic root restoration. The dataset contains 8,922 root images, including 1,206 soil-cultivated roots and 7,716 hydroponic roots.

Building on this, the team developed the Edge-UNet Generative Adversarial Network (EU-GAN) based on a GAN framework. By adding an Edge Attention Module (EAM) in the generator, optimizing the loss function, and incorporating post-processing methods, they increased the cotton-root-restoration rate from 17.35 percent to 35.75 percent.