Optimal weighted GAN and U-Net based segmentation for phenotypic trait estimation of crops using Taylor Coot algorithm

Abstract

Accurate and robust collection of plant phenotypic data offers theoretical as well as technical support to support the growth of crop science and to ensure ecological security, agricultural growth, and food security. Identifying phenotypic traits of crops refers to the detection of difference exists in plant features caused due to interaction of the environment and plant genetics. It is an important research discussed in plant breeding as it permits breeders to find a variety of crops with physical features, like stress resistance, and high yield. Manual measurement of phenotypic traits in the area is labor intensive and causes inaccurate results and these issues are resolved by developing a method based on the Taylor Coot algorithm for segmenting plant regions and biomass area to detect emergence counting and to estimate the biomass of crops. The process of counting emergence and estimating the biomass is performed in a parallel way using a Deep Residual Network (DRN) that is trained by developed optimization. The segmentation framework is done using Generative Adversarial Network (GAN) and U-Net to segment the plant regions and biomass area. For instance, the extraction of vegetation indices makes the process of biomass estimation to generate more optimal features using a deep learning model. Moreover, the proposed model obtains minimal Mean Absolute Difference (MAD), Standard Absolute Difference (SDAD), %Difference (%D) as 0.073, 0.074, and 16.45 for emergence counting. Moreover, the DRN shows higher performance by attaining minimum MAD, SDAD, and %D as 0.069, 0.096, and 14.85 for biomass estimation.