Retrieval of Tobacco Canopy Chlorophyll content by integrating multispectral vegetation indices and texture features

This is the authors' abstract. We don't add key points for this paper.

To achieve efficient and non-destructive monitoring of chlorophyll content in tobacco canopies, this study proposes an inversion model that integrates unmanned aerial vehicle (UAV)-based multispectral vegetation indices (VIs) and texture features (TFs). Based on six-band multispectral imagery acquired during the vigorous growth stage of tobacco, a set of VIs and TFs were constructed. The minimum Redundancy Maximum Relevance (mRMR) algorithm was employed to select the top 20 VIs and 20 TFs with the highest relevance. Subsequently, a Gradient Boosting Decision Tree (GBDT) combined with five-fold cross-validation was used to further select 9 key VIs and 12 representative TFs. On this basis, a dual-module model named MLP-GBDT, which integrates feature interpretation and decision optimization, was developed and compared with four models: Random Forest (RF), Support Vector Regression (SVR), GBDT, and Multilayer Perceptron (MLP). The results indicate that: (1) the integration of VIs and TFs significantly improved inversion accuracy. Compared to using single features alone, the coefficient of determination (R²) increased by 20.7%–49.8% and 3.8%–17.4%, respectively; (2) the proposed MLP-GBDT model achieved the best performance, with an R² of 0.851, root mean square error (RMSE) of 1.214, and mean absolute error (MAE) of 1.007 under the fused feature input, representing a 10.1%–16.7% improvement in R² compared to the other models. Overall, the multi-feature fusion-based MLP-GBDT model substantially enhances the accuracy of chlorophyll content inversion in tobacco leaves, offering both theoretical support and technical guidance for the application of UAV remote sensing in tobacco cultivation.