A novel inversion method of tobacco canopy nitrogen content based on vegetation index and color index

• Combined Color Indices (CCIs) have overcome single color feature limitations by integrating with Vegetation Index (VIs), significantly improving nitrogen inversion accuracy. • Variable Importance Projection (VIP) has demonstrated superior performance over Recursive Feature Elimination (RFE) in complex feature screening, optimizing model capability. • The fusion of Gradient Boosting Decision Tree (GBDT) and Support Vector Machine (SVM) model improves the accuracy and stability. • 12-band data, enriched with spectral information, has demonstrated superior performance compared to 6-band data. Aiming at the problem of insufficient accuracy of traditional single feature and traditional modeling methods in the inversion of tobacco canopy nitrogen content, this study proposes an inversion method combining combined feature and model fusion strategy. Based on the 6-band and 12-band multi-spectral image data, the vegetation index (VIs) and color feature (CF) were extracted, and the combined color index (CCIs) was further constructed. On this basis, the correlation between VIs, CF, CCIs and their fusion features and nitrogen content was comprehensively analyzed, and the most representative features were selected by variable importance projection (VIP) algorithm and recursive feature elimination (RFE) algorithm. In the modeling stage, a fusion model (GBDT-SVM) and two benchmark models (GBDT and SVM) were constructed to systematically compare the effects of different model structures on the accuracy of nitrogen content inversion. The results show that: (1) Under the same input variable conditions, the VIP method is superior to the RFE method as a whole, and the effect of screening features is more advantageous; (2) Regardless of the VIP or RFE algorithm, the performance of CCIs is better than that of CF. Overall, the combined feature input is significantly better than the single feature input, and the combined input of CCIs and VIs is the most stable. (3) Compared with the baseline model, the GBDT-SVM fusion model showed better performance under different feature selection methods and data combinations. Among them, R 2 exceeded 0.9, RMSE and MAE were significantly lower than those of the single model, and the prediction accuracy was significantly improved. (4) Compared with 6-band data, 12-band data performed better in nitrogen inversion accuracy and stability, especially the addition of red edge and near-infrared bands, which improved the performance of the model through richer spectral information and enhanced the sensitivity and prediction ability to nitrogen content changes. Through the feature and model fusion strategy, the inversion accuracy of tobacco canopy nitrogen can be effectively improved, which provides a new idea for tobacco leaf nitrogen management in precision agriculture.