Estimation of winter wheat LAI and SPAD based on the fusion of texture features and vegetation indices from UAV images

Accurate and efficient quantitative estimation of leaf area index (LAI) and soil and plant analyzer development (SPAD) in winter wheat is significant for field management decisions and yield prediction. This study compared the performance of three linear regression techniques: multiple linear regression (MLR), ridge regression (RR), and partial least squares regression (PLSR) and three machine learning algorithms: back-propagation neural networks (BP), random forests (RF), and support vector machine (SVM) with spectral vegetation indices (VIs), texture features (TEs), and their combinations extracted from UAV RGB images. A total of 36 estimation models were constructed, which included 24 models based on single datasets (VIs-based and TEs-based), and 12 models based on data fusion (VIs+TEs-based). The results revealed that combining VIs and TEs improved the accuracy of LAI and SPAD estimation for wheat compared to using VIs or TEs alone. Moreover, different data dimensionality reduction methods, including principal component analysis (PCA) and stepwise selection (ST), were used to improve the accuracy of LAI and SPAD estimation. The results showed that ST, PCA, and ST_PCA methods have different impacts on the accuracy of estimating crop parameters with the combination of VIs and TEs, where ST_PCA is effective in dealing with high-dimensional data and maintaining the accuracy of the model. The RF model combined with ST_PCA for integrating VIs and TEs achieved the best estimations, with R² of 0.86 and 0.91, RMSE of 0.26 and 2.01, and MAE of 0.22 and 1.66 for LAI and SPAD, respectively. ST_PCA, combined with machine learning algorithms, holds promising potential for monitoring crop physiological and biochemical parameters.