Abstract: To address the issues of low efficiency in the traditional measurement of tobacco leaf area index (LAI), feature redundancy in remote sensing inversion, and the tendency of model parameter optimization to easily fall into local optima, tobacco LAI estimation models suitable for different growth stages were constructed. The rosette stage and vigorous growth stage of tobacco were taken as the target growth stages, and 25 remote sensing variables were extracted based on unmanned aerial vehicle (UAV) multispectral imagery. Feature selection was conducted using competitive adaptive reweighted sampling (CARS), least angle regression (LARS), and iteratively retains informative variables (IRIV) algorithms. The results demonstrated that the IRIV algorithm exhibited superior comprehensive performance and could effectively reduce prediction errors. Based on the optimally selected features, Random Forest (RF), XGBoost, Back Propagation (BP) neural network, and Support Vector Regression (SVR) models were established, among which the SVR model demonstrated the best baseline predictive capability. Furthermore, a Particle Swarm Optimization algorithm improved by the Tent chaotic map (TPSO) was introduced to optimize the parameters of the SVR model, thereby constructing the TPSO-SVR model. Compared with the PSO-SVR model, the coefficient of determination (R²) of the TPSO-SVR model on the test set at the rosette stage increased by 0.015 to 0.757, and the root mean square error (RMSE) decreased by 0.006 to 0.243; at the vigorous growth stage, the R² on the test set increased by 0.021 to 0.872, and the RMSE decreased by 0.020 to 0.253. The results indicated that the stage-specific inversion method based on IRIV feature selection and the TPSO-SVR model could improve the estimation accuracy of tobacco LAI at key growth stages, providing a technical reference for the non-destructive monitoring of tobacco growth and precision field management.

Key words: tobacco, LAI, feature selection, chaotic sequence, PSO, Tent map