Quantifying effect of tassels on near-ground maize canopy RGB images using deep learning segmentation algorithm.

Vegetation indexes (VIs) are a key variable for monitoring the crop growth and estimating crop productivity. The spectral traits of the tassels were significantly different from those of the canopy leaves, which had an effect on the accuracy of using vegetation indices of canopy features to monitor crop growth. Advances in deep learning in computer vision provide opportunities for rapid and nondestructive tassels removal. The objective of this study was to remove tassels from images using a rapid and non-destructive method and to analyze and quantify the effect of tassels on canopy spectral information. This study proposed the color space transformation to divide tassels images into different growth stages and analyzed the performance of three convolutional neural networks (U-Net, PSPNet and DeeplabV3+) for tassel segmentation, and the analysis and quantification of the effect of tassels on canopy spectral information by calculating VIs and contributions. The results showed that the HSV (Hue, Saturation, Value) color space transformation divided images into three growth stages with robustness in the field environment. The U-Net model had the best segmentation precision (PA = 82.14%, MIoU = 84.43%). Moreover, the tassels significantly reduce canopy reflectance in the green wavelength range and the contribution of tassels to each of the VIs was different. In addition, the contribution of tassels to VIs is dynamic in different stages of tassels growth. This study provides a fast and non-destructive approach to quantify the effect of tassels on canopy spectral information in the field environment. [ABSTRACT FROM AUTHOR]