Rapid and accurate estimation of rice quality in northeast China based on UAV multispectral images

Objectives Accurate assessment of rice quality before harvesting is required for improving rice nutrient management and achieving both superior quality and market value.

Methods Two paddy fields, where soil fertility was different from each other, were selected for the research in Daxing and Qinglongshan farms of Jiamusi City, Heilongjiang Province. Soil samples were collected in every 20 m×20 m area for determination of available N content, then nitrogen fertilizer amounts were calculated for reaching five target rice yield levels, and the distinct rice populations were created through machine interpolation and variable nitrogen application. The DJI Phantom 4 quadrotor multispectral UAV was used to obtain the canopy multi-spectral data of rice at tillering, jointing, heading and maturity stages, respectively, and the data were transferred into five type of normalized vegetation indexes. At maturity, rice grain samples were collected from notably different rice populations for determination of protein, amylose content, yield, and taste quality, and corresponding soil samples were collected afterwards for analysis of organic matter, available N, P, and K contents. 67% of the acquired data at each growth stage were used to construct the multiple linear regression models between the quality indices and each of the vegetation index, using the determination coefficient (R²) and the square root of the variance (RMSE) to assess the accuracy of estimated quality. And the left 33% of data were used for the testify of the assessment.

Results At maturity stage of rice, the coefficients of variation (CV) of soil organic matter, available N, P and K contents were 11.65%, 14.44%, 37.66% and 11.60% in Daxing farm, and the CVs were 14.45%, 14.32%, 36.37% and 28.51% in Qinglongshan farm, respectively. At maturity, the CVs of rice yield and taste value were all >10% in the two sites, the CV of protein was >10% in Qinglongshan, while the CVs of amylose varied in range of 1.11%−1.83% in the two sites, indicating that the taste value and protein content of rice were suitable, while the amylose content was unsuitable for further quality assessment. From tillering to heading stages, the R² and RMSE of multiple regression models were 0.262−0.794 and 0.259−0.686 for protein content, and 0.240−0.755 and 4.211−7.588 for taste value; and at maturity stage, the R² and RMSE were 0.791−0.874 and 0.166−0.365 for protein content, and 0.786−0.852 and 2.836−4.039 for taste value, respectively. The estimation accuracy of protein and taste values based on vegetation index during the mature stage was better than the heading stage. Incorporating soil indicators such as pH, organic matter, available N, P and K into multiple regression models at the heading stage enhanced the estimation accuracy significantly, the R² of the protein model in Daxing farm increased from 0.585 to 0.720, and the RMSE decreased from 0.301 to 0.247. The R² of the taste value models in two farms increased from 0.565−0.755 to 0.706−0.787, while the RMSE decreased from 4.318−4.854 to 3.993−4.029. However, the combination of soil fertility and vegetation index during the mature stage cannot improve the prediction accuracy of the model.

Conclusions Assessment of rice quality is feasible by multispectral UAV canopy images at maturing stage of rice. The assessment using the images at heading stage should include the soil nutrient contents to enhance the estimating accuracy.