基于光谱和纹理信息空间尺度优化的夏玉米冠层EWT反演模型

EWT Inversion Model of Summer Maize Based on Spatial Scale Optimization of Spectral and Texture Information

摘要:

针对空间异质性导致的冠层等效水厚度(Equivalent water thickness, EWT)反演误差较大的问题，以4块长势差异较大的玉米田为研究对象，分别采集6个关键生育节点的EWT数据，同时利用无人机多光谱遥感技术获取田间的正射影像。以滑动窗口的方式提取遥感影像不同窗口空间尺寸(0.1m×0.1m~2.0m×2.0m)的光谱和纹理信息，经多重共线性检验后，应用主成分分析法(Principal component analysis，PCA)分别对光谱参数(Spectral parameters，S)、纹理参数(Texture parameters，T)及光谱与纹理组合参数(Spectral and texture parameters，S+T)进行降维，进而分别利用偏最小二乘法(Partial least squares，PLS)、随机森林(Random forest，RF)以及支持向量机(Support vector machine，SVM)构建EWT反演模型，而后利用Kruskal-Wallis检验模型的精度，并根据多重检验结果探讨最佳窗口尺寸的选择。结果表明：随着窗口空间尺度的逐渐增大，EWT反演模型的精度呈先增大后减小趋势；以S+T作为输入参数构建的模型精度显著优于S和T，引入纹理特征后，基于PLS、RF和SVM的模型最优窗口尺寸校正决定系数(Adjusted R-square，R2adj)分别增加0.16、0.05和0.12，相对均方根误差(Relative root mean square error，RRMSE)分别减小4.95%、1.17%和3.80%，表明纹理特征可以提高EWT模型反演精度；综合比较不同建模方法构建的9组模型，确定最优采样窗口空间尺寸为0.7m×0.7m(R2adj最高可达0.82，对应的RRMSE为16.57%)。该研究可为基于无人机多光谱影像分析的信息挖掘和EWT监测提供参考。

Abstract:

In order to solve the problem of large canopy equivalent water thickness (EWT) inversion error caused by spatial heterogeneity, taking four maize fields with large growth differences as the research object, EWT data of six key growth nodes was collected, and UAV multispectral remote sensing technology was used to obtain orthophoto images in the field, and the spectral and texture information of different window space sizes (0.1m×0.1m to 2.0m×2.0m) of remote sensing images in the form of sliding windows was extracted, and after multicollinearity testing, principal component analysis (PCA) was used to reduce the dimensionality of spectral parameters (S), texture parameters (T) and combinatorial parameters (S+T), respectively, and then the EWT inversion model was constructed by partial least squares (PLS), random forest (RF) and support vector machine (SVM), respectively, and then the accuracy of the model was tested by Kruskal-Wallis, and the choice of optimal window size was discussed according to the results of multiple tests. The results showed that with the gradual increase of the window space scale, the accuracy of the EWT inversion model was increased firstly and then decreased. The accuracy of the model constructed with the S+T as the input variable was significantly better than that of the S and the T, and the adjusted R-square (R2adj) of the optimal window size of the model based on PLS, RF and SVM was increased by 0.16, 0.05 and 0.12, respectively, and the relative root mean square error (RRMSE) was decreased by 4.95%, 1.17% and 3.80%, respectively. The results showed that it was feasible to use texture features to improve the inversion accuracy of EWT model. Comprehensively comparing the nine sets of models constructed by different modeling methods, the optimal sampling window spatial size was finally determined to be 0.7m×0.7m, with R2adj up to 0.82 (corresponding RRMSE of 16.57%). The research result can provide a reference for information mining and EWT monitoring based on UAV multi-spectral image analysis.