融合无人机光谱信息与纹理特征的棉花叶面积指数估测

农田尺度下作物叶面积指数(Leaf area index, LAI)的精准监测，对于研究群体结构对产量和管理措施的响应具有重要意义。目前普遍采用无人机光谱特征反演作物的LAI指数，作为长势和冠层结构诊断的重要依据，其估测精度的准确性是否可以提高仍有待研究。作物表面特征，如灰度和颜色，在不同生育阶段会发生变化。为此，本研究考虑到LAI的影响因素，设置不同的种植密度和氮素水平营造差异化的冠层结构，利用搭载多光谱传感器的无人机获取主要生育时期棉花的冠层图像得到植被指数(Vegetation indexs, VIs),基于二阶概率统计滤波(Co-occurrence measures)方法获取均值(MEA)、方差(VAR)、协同性(HOM)、对比度(CON)、相异性(DIS)、信息熵(ENT)、二阶矩(SEM)和相关性(COR)等8个纹理特征值(Texture features, TFs)。最后，采用支持向量机回归(SVR)、偏最小二乘法(PLSR)、深度神经网络(DNN)分别建立基于光谱特征、纹理特征以及二者结合的棉花LAI的估算模型，并比较差异。试验结果表明：VI₍(...

Cotton Leaf Area Index Estimation Combining UAV Spectral and Textural Features

Accurate prediction of crop leaf area index (LAI) at farm scale is important for studying the response of population structure to yield and management practices. The inversion of the LAI of crops by spectral features from drones is now commonly used as an important basis for diagnosing crop growth and canopy structure, and it remains to be investigated whether the accuracy of its estimation can be improved. Crop surface features, such as greyscale and colour, can change under different levels of structural complexity. For this reason, the influence of LAI was taken into account by setting different planting densities and nitrogen levels to create a differentiated canopy structure, using an unmanned aerial vehicle with a multispectral sensor to obtain canopy images of cotton during the main fertility periods to obtain vegetation indices and second-order probability-based statistical filtering (co-occurrence measures) in the near infrared band to extract mean (MEA), variance (VAR), synergy (HOM), contrast (CON), dissimilarity (DIS), information entropy (ENT), second order moments (SEM) and correlation (COR) of the eight texture feature values. Finally, support vector regression (SVR), partial least squares regression (PLSR) and deep neural networks (DNN) were used to develop models for estimating cotton LAI based on spectral features, texture features and a combination of the two, respectively, and to compare the differences. The results showed that the vegetation indices VI(nir/green), VI(nir/red), GNDVI, OSAVI and mean had high correlation with LAI; the LAI estimation accuracy established by SVR was the highest (R2=0.78, RMSE was 0.22, RRMSE was 0.10); among three estimation models, the SVR model combining VIs and texture features improved the accuracy by 7.89% (VIs) and 32.26% (TFs), respectively, over the single parameter type model. Thus the LAI estimation model incorporating UAV spectral information and image texture provided a feasible and accurate method for the diagnosis of cotton canopy structure in dense crops.