基于无人机RGB图像植被指数的棉花产量估算研究

Cotton Yield Estimation Based on UAV RGB Image-based Vegetation Indexes

摘要:

为突破传统遥感技术依赖多光谱/高光谱的局限，探索RGB图像结合颜色空间转换算法的作物产量估算方法。通过无人机平台获取16个水氮处理条件下滴灌棉花6个关键生育期的冠层RGB图像，运用色彩空间转换算法将其转化为HIS、CIELab和CIELuv颜色参数，优选最佳估产窗口期，基于其衍生的RGB植被指数构建3种变量组合下的岭回归、支持向量机和随机森林估产模型。不同生育期的RGB植被指数与产量相关性表明，RGB植被指数与棉花产量在花期、花铃Ⅰ期、花铃Ⅱ期、铃期和吐絮期均具有较强的相关性，吐絮期植被指数与产量相关性最强，其中吐絮期的估产精度最高（决定系数大于等于087，偏差小于10%），为最佳估产窗口期。随机森林模型在各生育期估产中反演精度表现最稳定，采用变量组合3（GA、GGA、CSI、NGRDI、NGRDIveg、TGI、TGIveg、NDLab、NDLuv）构建的随机森林模型反演结果表现最优，测试集决定系数为076~088，均方根误差为0.69~0.99t/hm2，平均绝对误差为0.53~0.80t/hm2，偏差为6.11%~30.65%，为滴灌条件下棉花产量最优反演模型。研究结果可为利用无人机RGB图像进行滴灌棉花估产以及表型监测分析提供理论参考。

Abstract:

In order to overcome the limitations of conventional remote sensing technologies that depend on multispectral or hyperspectral imaging, the potential of integrating RGB images with a color space conversion algorithm was explored for the purpose of crop yield estimation and monitoring. The canopy RGB images of drip-irrigated cotton acquired via UAV at six growth stages under 16 distinct water and nitrogen treatments were utilized. These images were converted into HIS, CIELab, and CIELuv color parameters through the implementation of a color space conversion algorithm. The most suitable yield estimation window was identified through a systematic selection process. Based on the derived RGB vegetation index, three machine learning algorithms, including ridge regression, support vector machine, and random forest were used to construct the drip-irrigated cotton yield in different growth stages under three different variable combinations. The findings demonstrated a robust correlation between the RGB vegetation index and cotton yield during various growth periods. The correlation was particularly pronounced in the flowering stage, flowering and boll stage Ⅰ, flowering and boll stage Ⅱ, boll-setting stag, and boll opening stage. The correlation between vegetation index and yield in the boll opening stage exhibited the strongest correlation, and the yield estimation accuracy in the boll opening stage was the highest (coefficient of determination was greater and equal to 087, deviation was less than 10%). The boll opening stage window demonstrated the optimal yield estimation. The inversion accuracy of the random forest machine model exhibited the most optimal comprehensive performance. The inversion result of the random forest model constructed by variable combination 3 (GA, GGA, CSI, NGRDI, NGRDIveg, TGI, TGIveg, NDLab, NDLuv) was the most optimal, with a test set determination coefficient of 0.76~0.88, a root mean square error of 0.69~0.99t/hm2, a mean absolute error of 0.53~0.80t/hm2, and a deviation of 6.11%~30.65%, which was the optimal inversion model for cotton yield under drip irrigation. The findings can serve as a theoretical foundation for the utilization of drone RGB images in the estimation of cotton yield from drip irrigation and the monitoring and analysis of phenotype.