Construction of a Prediction Model for Functional Traits of Grape Leaves Based on Multi-Stage Collaborative Optimization

The efficient detection of grape leaf nutrient parameters, including chlorophyll content, represented by soil and plant analysis development (SPAD), leaf nitrogen content (LNC), leaf potassium content (LKC), fresh weight water content (FWC), and dry weight water content (DWC), is crucial in precision agriculture. This study introduces a modeling framework that integrates hyperspectral preprocessing, feature selection, and multimodal data fusion. This framework enhances feature representation and model robustness by fusing spectral features (Ref), vegetation indices (VIS), and color and texture features from hyperspectral and red, green, and blue (RGB) images. Comparative experiments based on partial least squares regression (PLSR), Gaussian Process regression (GPR), and Bayesian Ridge regression (BRR) demonstrate that with a limited sample size, the PLSR and BRR models exhibit superior predictive performance and stability. However, during the optimization process, the performance improvement of the GPR model was the greatest (with R2 increasing by up to 31.9%). Among the features, vegetation indices showed relatively high correlations with various traits. For image features, hyperspectral texture characteristics performed best, while color features from RGB images contributed significantly. Following preprocessing, feature selection, and feature combination, the performance of all models, except for DWC, improved progressively. Notably, feature selection significantly increased model accuracy. These findings indicate that multi-stage collaborative optimization strategies can be employed for the precise prediction of grape leaf functional traits.