融合无人机光谱信息与纹理特征的冬小麦综合长势监测

Comprehensive Growth Monitoring of Winter Wheat by Integrating UAV Spectral Information and Texture Features

摘要:

高效、及时获取作物长势信息对作物生产管理具有重要作用。目前针对小区域农作物长势监测多以无人机光谱信息反演来实现，但综合考虑农作物不同生育期阶段的表面特征信息进行小区域农作物长势监测的方法需进一步研究。本文以冬小麦为研究对象，基于冬小麦株高和叶面积指数（Leaf area index，LAI）按照变异系数法构建综合长势监测指标（Comprehensive growth monitoring indicators，CGMI），提出一种融合无人机光谱信息与纹理特征的冬小麦综合长势监测方法。以搭载多光谱镜头的无人机获取冬小麦4个生育期影像，得到12种植被指数和各波段的8类纹理特征。采用Person相关性分析方法，筛选出与CGMI相关性较好的植被指数与纹理特征，进而采用随机森林回归（Random forest，RF）、偏最小二乘回归（Partial least squares regression，PLSR）、支持向量机回归（Support vector regression，SVR）3种机器学习方法分别构建基于植被指数和基于植被指数与纹理特征的2个长势监测模型，通过比较得到较优长势监测模型，最终获得研究区冬小麦长势空间分布信息。结果表明：3种机器学习方法中，基于植被指数与纹理特征的SVR长势监测模型精度最高（训练集R2为0.789，MAE为0.03，NRMSE为4.8%，RMSE为0.04），与基于植被指数的SVR长势监测模型相比，该模型决定系数提高5.1%，平均绝对误差降低3.3%，标准均方根误差降低8.3%，均方根误差降低10%。研究结果证明该方法精确、可靠，可为冬小麦长势监测提供参考。

Abstract:

Efficient and timely acquisition of crop growth information plays an important role in crop production management. At present, crop growth monitoring in small areas is mostly achieved through the inversion of spectral information from UAV. However, further research is needed to comprehensively consider the surface feature information of crops at different growth stages for monitoring crop growth in small areas. Taking winter wheat as the research object, comprehensive growth monitoring indicators (CGMI) was constructed based on the plant height and leaf area index (LAI) of winter wheat according to the coefficient of variation method, and a comprehensive growth monitoring method was proposed for winter wheat that combining UAV spectral information and texture features. A drone equipped with a multispectral lens was used to acquire images of winter wheat in four growth stages, and 12 vegetation indices and 8 types of texture features in each band were obtained. The Person correlation analysis method was used to screen the vegetation index and texture features that had good correlation with CGMI, and then random forest regression (RF), partial least squares regression (PLSR) and support vector regression (SVR) methods were used to construct growth monitoring models based on vegetation index and growth monitoring models based on vegetation index and texture features, respectively. Through comparison, the superior growth monitoring model was obtained, and finally the spatial distribution information of winter wheat growth in the study area was obtained. The results showed that among the three machine learning methods, the SVR growth monitoring model based on vegetation index and texture features had the highest accuracy (training set R2 was 0.789, MAE was 0.03, NRMSE was 4.8%, RMSE was 0.04). Compared with SVR growth monitoring model based on vegetation index, the coefficient of determination of this model was increased by 5.1%, the average absolute error was decreased by 3.3%, the standard root mean square error was decreased by 8.3%, and the root mean square error was decreased by 10%. The research results showed that the method was accurate and reliable, which can provide an important reference for winter wheat growth monitoring.