基于无人机图像分割的冬小麦叶绿素与叶面积指数反演

叶绿素含量与叶面积指数是反映作物长势的重要理化参数,准确、高效定量估计小麦叶绿素含量与叶面积指数对于产量预测和田间管理决策具有重要意义,无人机(Unmanned Aerial Vehicle,UAV)遥感影像具有高空间分辨率的优势,被广泛应用于作物理化参数反演,但现有叶绿素含量与叶面积指数反演模型受土壤、阴影等背景噪声...

Inversion of chlorophyll and leaf area index for winter wheat based on UAV image segmentation

Unmanned aerial vehicle (UAV) remote sensing images with a high spatial resolution have been widely used to estimate crop physical and chemical parameters in recent years. Since the planting area of wheat has accounted for about 1/5 of the total grain one in China, an accurate and efficient estimation of wheat phenotypic parameters is of great significance for yield prediction and field management. Leaf Area Index (LAI) and chlorophyll relative content are closely related to the ability of crops to intercept and absorb the incident Photosynthetically Active Radiation (PAR). These key variables have been excellent indicators for various abiotic and biological stresses in the photosynthesis, respiration, and transpiration during crops growth. However, the existing estimation models are greatly affected by some background noises in the remote sensing images, such as soil and shadow. The objective of this study was to explore whether the removal of background pixels was utilized to improve the inversion accuracy of chlorophyll content and LAI of crops. The Excess Green minus Excess Red vegetation index (ExG-ExR) was first selected to segment the UAV multi-spectral images of winter wheat in the key growth periods (jointing, flagging, and flowering). Then, the average reflectance of wheat pixels (GreenPix) was extracted in the test plot. Five multispectral bands and 20 vegetation indices were selected to analyze the correlation with three phenotypic parameters, including the Soil and Plant Analysis Development (SPAD), LAI, and Canopy Chlorophyll Content (CCC). The first five sensitive vegetation indices with the high correlation were screened to establish the inversion models of physical and chemical parameters in winter wheat using Partial Least Squares Regression (PLSR). The results showed that the model of wheat pixel spectrum (VI_GreenPix) was significantly improved the estimation accuracy of winter wheat SPAD (Calibration dataset: R2 = 0.85, RMSE = 3.51; Verification dataset: R2=0.93, RMSE=2.67), indicating a higher estimation accuracy of SPAD, compared with all pixel spectrum (VI_AllPix) (Calibration dataset: R2=0.79, RMSE=4.12; Verification dataset: R2 = 0.78, RMSE = 4.19) under the whole coverage, especially the coverage below 40%. The accuracy of LAI inversion by VI_GreenPix was consistent with the constructed model by VI_AllPix (R2=0.70, RMSE = 0.42), but the verification accuracy was much higher (R2=0.80, RMSE=0.38) than that by VI_AllPix (R2=0.75, RMSE=0.42). The VI_GreenPix was greatly contributed to the LAI estimation accuracy, when the coverage was less than 80%. The VI_GreenPix also improved the inversion accuracy of CCC under the whole coverage (Calibration dataset: R2 =0.79, RMSE=21.14; Verification dataset: R2= 0.69, RMSE=23.50), with the best effect at the coverage less than 70%. Consequently, the spectral information of vegetation pixels was extracted to construct the physical and chemical parameters estimation model of winter wheat, further to improve the inversion accuracy of SPAD, LAI, and CCC via the vegetation index threshold segmentation using high-resolution UAV multispectral images. The findings can provide a technical support to the growth monitoring and yield prediction of winter wheat.