基于多源无人机影像特征融合的冬小麦LAI估算

为探讨无人机多源影像特征融合估测作物叶面积指数的能力，该研究以冬小麦为研究对象，利用多旋翼无人机搭载高清数码相机和UHD185成像光谱仪获取研究区冬小麦关键生育期（扬花期、灌浆期）的可见光和高光谱影像。综合考虑可见光、高光谱影像特征与冬小麦叶面积指数的相关性及影像特征重要性进行特征筛选，然后，以可见光植被指数、纹理特征、可见光植被指数+纹理特征、高光谱波段、高光谱植被指数及高光谱波段+植被指数分别作为输入变量构建多元线性回归、支持向量回归和随机森林回归的叶面积指数估测模型（单传感器数据源）；以优选的两种影像特征结合支持向量回归、随机森林回归构建叶面积指数估测模型（两种传感器数据源），比较分析单源与多源影像特征监测冬小麦叶面积指数的性能。进一步地，考虑到小区土壤空间异质性会影响冬小麦叶面积指数估测结果，该研究探讨了不同影像采样面积下基于单源遥感数据构建的小麦叶面积指数估测模型精度。研究结果表明：在扬花期和灌浆期，使用两种影像优选特征构建的随机森林回归估测模型精度最佳，验证集决定系数分别为0.733和0.929，均方根误差为0.193和0.118。可见光影像采样面积分别为30%和50%，高光谱影像采样面积为65%时，基于单源影像特征构建的随机森林回归估测模型在扬花期和灌浆期效果最好。综上，该研究结果可为无人机遥感监测作物生理参数提供有价值的依据和参考。

Estimation of winter wheat leaf area index using multi-source UAV image feature fusion

Leaf area index (LAI) is a key indicator for the growth of crops. The crop yield is also closely related to the LAI, particularly for the decision-making in modern agriculture. Rapid and accurate detection of the crop LAI is of great significance to field production management. Single sensors have been mostly used to monitor the winter wheat LAI in the past, such as high-definition digital cameras, multispectral and hyperspectral cameras. Fortunately, Unmanned Aerial Vehicles (UAV) remote sensing technology has been developed for crop LAI monitoring by virtue of the high timeliness and low cost at present. The multi-source image data can also be combined for parameter monitoring. In addition, previous LAI estimation is limited to either only the correlation of image features with LAI, or only the importance of image features. Therefore, this study aims to comprehensively consider the correlation of image feature with the LAI and image feature importance, and then construct the multi-source remote sensing LAI estimation models using the selection of optimal image features. Baihu Farm in Lujiang County and Shucheng County Agricultural Science Institute in Anhui Province of China were selected as the study areas, where the canopy visible and hyperspectral images of winter wheat at flowering and filling stages were collected by a UAV platform equipped with a high-definition digital camera and an imaging hyperspectrometer. Meanwhile, the ground LAI data was collected using LAI-2200C. The multiple linear regression (MLR), support vector regression (SVR), and random forest regression (RFR) algorithms were selected to estimate the wheat LAI using the visible and hyperspectral image data. Firstly, the correlation between the visible and hyperspectral image features with the winter wheat LAI was analyzed as well as the importance of image features were calculated, where the optimal image features were selected. Secondly, the MLR, SVR, and RFR estimation models (single-sensor data sources) were constructed, where the input was taken as the visible vegetation index, the texture features, visible vegetation index combined with texture features, hyperspectral band, hyperspectral vegetation index, and hyperspectral band combined with vegetation index. The RFR and SVR LAI estimation models (two sensor data sources) were constructed with two image-preferred features to compare the performance of single-source and multi-source image features for the monitoring of wheat LAI. Thirdly, the spatial heterogeneity of plot soils was considered for the wheat LAI monitoring. The wheat LAI estimation model was also constructed to combine the single image features under different image sampling areas. The results showed that the best accuracy of the RFR LAI estimation model was achieved at the flowering and filling stages using two image preferred features, with the validation set R2 of 0.733 and 0.929 and RMSE of 0.193 and 0.118, respectively. By contrast, the model using the combination of single image features performed the best at the flowering and filling stages, when the sampling areas of visible images were 30% and 50%, and the sampling areas of hyperspectral images were 65%, respectively. In summary, the study can provide a valuable reference for the UAV remote sensing monitoring of crop physiological parameters.