GlandSegNet: Semantic segmentation model and area detection method for cotton leaf pigment glands

• Published in: Computers and electronics in agriculture 2023-09, Vol.212, p.108130, Article 108130
• Main Authors: Xu, Yaxuan, Wang, Guoning, Shao, Limin, Wang, Nan, She, Lixuan, Liu, Yu, Geng, Yuhong, Yan, Geng
• Format: Article
• Language: English
• Subjects: Area detection; Cotton leaf; Gossypol; Pigment gland; Semantic segmentation
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2023.108130

•Accurate segmentation of cotton leaves and pigment glands is realized using semantic segmentation.•A network model is constructed by first upsampling and then pooling and embedding an attention mechanism.•The proposed method enables rapid detection of pigment gland area information in the whole leaf.•Guidance for semantic segmentation techniques targeting small objects in organisms is provided. Pigment glands in cotton store gossypol, which is a highly valuable substance in agriculture and medicine. Therefore, obtaining phenotype information on cotton pigment glands is essential for evaluating gossypol content. However, the current research on pigment glands in cotton leaves faces several challenges, including a small proportion of pigment glands in the leaf area, a small segmentation target, a large number of glands at the leaf edge, and the interference of leaf veins, making the whole-leaf phenotype detection difficult. To address these challenges, this study proposes a deep learning-based semantic segmentation model named GlandSegNet. The GlandSegNet uses a specific encoder architecture formed by fusing the upsampling process and feature-extraction networks. In the decoder part, pooling and skip-connection operations are performed, and the network is optimized by embedding the ECA attention module. The experimental results show that the GlandSegNet can achieve area accuracy rates of 0.9842 and 0.9510, with corresponding error areas of 0.6966 mm2 and 4.1258 mm2, on test set 1 consisting of intact leaves from a single species and test set 2 consisting of randomly selected leaves from multiple species, respectively. The results demonstrate that the GlandSegNet semantic segmentation method can exhibit excellent performance in the cotton leaf pigment gland area detection tasks. Compared with traditional microscopic observation methods, the GlandSegNet is characterized by high efficiency and the ability to quantitatively analyze the glandular phenotype. Thus, the GlandSegNet could provide an effective tool and technical support for large-scale cotton pigment gland research and could be of great significance for the phenotype evaluation of cotton pigment glands and gossypol content evaluation.