Abstract: To accurately estimate tobacco leaf area index (LAI), multispectral images, light detection and ranging (LiDAR) data and ground LAI data were synchronously collected at three key times selected at the fast growing stage (T1), bud stage (T2) and early maturity stage (T3), respectively, in a tobacco experimental site in Xuanzhou District, Xuancheng City, Anhui Province. Tobacco vegetation indexes (VIs) and texture features (TFs) were extracted from the multispectral images, and the canopy height model (CHM) was derived from the LiDAR data. Feature selection was performed via feature importance analysis. Tobacco LAI inversion models at different sampling times were established based on back propagation neural network (BPNN), random forest regression (RFR), support vector regression (SVR), convolutional neural network (CNN) and long short-term memory (LSTM) network. The accuracy of these models was evaluated to identify the optimal LAI inversion model and the results showed that: 1) Compared with TFs, VIs had a stronger correlation with tobacco LAI. Among the single-feature models, VIs-based models generally outperformed TFs-based models, and the combination of VIs and TFs further improved the inversion accuracy. 2) CHM data effectively enhanced the LAI inversion accuracy, and the improvement effect was highly coupled with the correlation between LAI and CHM. Specifically, significant improvement was observed at T1 and T2, while only negligible improvement was found at T3. 3) The CNN model achieved the best LAI inversion performance with the combined input of CHM, VIs and TFs, yielding coefficients of determination (R²) of 0.822, 0.847 and 0.820 at the three sampling time points, respectively. This study provides a methodological reference for tobacco growth monitoring and yield estimation based on multi-source remote sensing data.