结合无人机光谱与纹理特征和覆盖度的水稻叶面积指数估算

为了探究无人机多指标构建叶面积指数(Leaf Area Index,LAI)估算模型的能力,该研究通过不同纹理组合方式优选纹理指数,分别以光谱特征、纹理指数和作物覆盖度作为输入量建立一元线性模型,3类指标结合构建多元逐步回归和人工神经网络模型,分析多指标结合估算LAI的精度.结果表明:新的纹理指数能够明显提高纹理特征值...

Estimation of rice leaf area index combining UAV spectrum, texture features and vegetation coverage

Paddy rice as an important food crop is exactly determining the national food security in China. Leaf Area Index (LAI) is then an important indicator to evaluate crop growth and field management. Dynamic information of rice growth can be gained from the LAI with the accumulation of aboveground biomass and yield formation. The unmanned aerial vehicles (UAV)-based multispectral remote sensing technology can quickly capture the information on spatial variability of crops at the field scale, due mainly to its higher temporal and spatial resolution. The differences in rice growth can therefore be gained within the plots. As such, the vegetation indices can be used to estimate crop LAI. But there are still some saturated limitations when the LAI is large in estimating LAI. In this study, a rice LAI estimation model was constructed to investigate the ability of UAV with multiple indicators, combining spectral features, texture indices,and crop coverage. The UAV multispectral images were used to extract the spectral information, texture features, and crop coverage. A combination of different texture features, including the difference, ratio, and normalization, was calculated to obtain new texture indices, and further to improve the correlation between texture features and LAI. A one-dimensional linear model was built, where the spectral features, the texture index, and crop coverage were used as input quantities. Three types of indicators were integrated to construct a multiple stepwise regression and artificial neural network model, where the accuracy of combining multiple indicators was analyzed to estimate LAI. K-fold cross-validation was adopted to verify the present model. The results showed that there were significant correlations between six vegetation indices and rice LAI. All correlation coefficients were above 0.6 and ranked in a descending order, the Optimized Soil-Adjusted Vegetation index (OSAVI), Modified Triangular Vegetation Index 2 (MTVI2), Difference Vegetation Index (DVI), Green Normalized Difference Vegetation Index (GNDVI), Normalized Difference Vegetation Index (NDVI), and red-edge Chlorophyll Index (CIRE). The combined texture features showed that the correlation coefficient of a single texture feature with the highest correlation was 0.731 before the operation, while the texture index significantly improved the correlation between texture feature values and LAI. Specifically, the mean combination of Normalized Difference Texture Index (NDTI), Difference Texture Index (DTI), and Ratio Texture Index (RTI) presented a high correlation with LAI, where the DTI (mean5, mean1) between the near-infrared band mean and the blue band mean was the highest correlation of 0.830, 13.54% higher than that the near-infrared band mean of a single texture feature. The highest accuracy was gained in the differential texture index and crop coverage combining GNDVI, when estimating the rice LAI. The multiple stepwise regression model combining multiple indicators (R2 =0.866, R2adj=0.816, RMSE=0.308) was significantly higher than that of a single vegetation index (R2=0.603, R2adj=0.563, RMSE=0.541), crop coverage (R2=0.633, R2adj=0.596, RMSE=0.516) and the LAI model constructed with a single texture index (R2=0.668, R2adj=0.635, RMSE=0.447). Better accuracy and some advantages of inversion were achieved to combine the spectral features, texture index, and crop coverage. The finding can provide a theoretical basis to estimate the structural parameters for the LAI of crops using the UAV platform in digital agriculture.