In-field chlorophyll estimation based on hyperspectral images segmentation and pixel-wise spectra clustering of wheat canopy

Abstract

Hyperspectral imaging (HSI) technology provides an efficient and non-destructive tool for studying leaf chlorophyll content (LCC) to evaluate photosynthesis capability and crop growth status. The challenges from soil background interferences and canopy complexities always reduce the estimation accuracy in field wheat. Thus, this paper aimed to improve accuracy of wheat LCC estimation by overcoming challenges from soil background and complex canopy. A method was proposed based on hyperspectral image segmentations and pixel-wise spectra clustering of wheat canopy. According to the spectral features of green vegetation, canopy segmentation methods were discussed based on near-infrared (NIR) image at 807 nm, excess green (ExG) image in visible bands (470, 550, 660 nm), and soil-adjusted vegetation index (SAVI) image calculated by 807 and 660 nm. Among the average spectra from raw HSI regions, three sequence spectra were from the segmented canopy based on NIR, ExG, and SAVI, the modelling results of RP2 presented an ascending tendency following 0.66, 0.70, 0.79, and 0.81, and RMSEP exhibited a gradually descending tendency of 5.24, 4.11, 3.65, and 3.14 mg L−1. These findings suggest that SAVI segmentation gained the best estimation result with the optimal soil background removal performance. After canopy segmentation, K-means clustering was applied to divide pixel-wise spectra of wheat canopy into four clusters. Partial least squares (PLS) regression results of the four clusters showed RP2 of 0.81, 0.86, 0.89, and 0.83 and RMSEP of 3.55, 2.97, 2.64, and 3.79 mg L−1, respectively. The spectra clusters showed the potential to reduce the influences of canopy complexities. As a result, the third pixel-wise spectra cluster based on SAVI image segmentation contributed to the improvement of the LCC estimation.