Full growth period inversion of peanut canopy chlorophyll content based on UAV multispectral data and machine learning: A stage-specific optimization strategy

Accurate monitoring of crop canopy chlorophyll content (CCC) is of significance for precision agriculture and sustainable development. However, current remote sensing retrieval of crop chlorophyll is often limited to a single growth stage, overlooking dynamic changes in canopy structure from seedling to maturity, prevents the dynamic monitoring of peanut CCC throughout the entire stage. Therefore, this study acquired UAV multispectral imagery at four key growth stages of peanut and extracted a feature set including 54 VIs, 32 TFs, and 64 DWT features. The LassoLarsCVFS algorithm was employed to identify the optimal feature subset for each growth stage. Subsequently, retrieval models were developed for using Lasso, ENR, SVR, and MLP. The results showed that the combination of VIs and the MLP model achieved the best performance at the seedling stage (R² = 0.588). At the flowering and pinning stage, the VIs+TFs feature set combined with the SVR model yielded the optimal result (R² = 0.676). At the podding stage, the VIs+DWT feature set with the SVR model performed best (R² = 0.733). At the maturity stage, the VIs+TFs feature set paired with the MLP model produced the highest accuracy (R² = 0.828). Compared with the global model, the proposed stage-specific strategy demonstrated superior performance, achieving improvements in R² of 36.7 %, 0.7 %, 1.8 %, and 6.6 % at the seedling, flowering and pinning, podding, and maturity stages, respectively. These findings indicate that dynamic feature selection and model optimization tailored to different growth stages are crucial for retrieving crop chlorophyll content throughout the entire growth stage. The stage-wise optimization strategy, by integrating spectral and multi-scale structural information, offers an effective approach to overcome the challenges of remote sensing retrieval in dense canopies and provides reliable support for precision nutrient management in peanut and other crops. • Fusing multiple features from UAV multispectral imagery significantly enhances the retrieval accuracy of peanut CCC. • Optimal combinations of remote sensing features and machine learning models are demonstrated to be stage-specific. • The LassoLarsCVFS proves to be an effective method for selecting salient features across different peanut growth stages. • Non-linear models, particularly SVR and MLP, exhibit superior performance for crop growth monitoring.