Leaf Area Index Inversion of Cotton Based on Drone Multi-Spectral and Multiple Growth Stages

The leaf area index (LAI) is a vital indicator for evaluating crop growth, photosynthesis, and transpiration. The objective of this study is to explore the cotton LAI estimation models based on multi-spectral data from drones at different growth stages and multiple growth stages, clarify the variation patterns of cotton LAI estimation models during different growth stages, and to provide a basis for real-time understanding of cotton growth and scientific field management tailored to local conditions.

The DJI Elf 4 multi-spectral UAV was used to acquire multi-spectral images and RGB images of cotton at budding stage, initial flowering stage, boll setting and open-boll stages. Five multi-spectral indices, namely normalized difference vegetation index (NDVI), normalized green difference vegetation index (GNDVI), normalized difference red-edge index (NDRE), leaf chlorophyll index (LCI), optimized soil adjusted vegetation index (OSAVI), and five color indices, namely modified green-red vegetation index (MGRVI), green-red vegetation index (GRVI), green leaf algorithm (GLA), excess red index (EXR), and visible atmospherically resistant vegetation index (VARI), were selected to build a data set for each growth stage of cotton and multiple growth stages of cotton growth, respectively. Combined with the punching method to obtain actual ground LAI data, the machine learning algorithms of partial least squares regression (PLSR), ridge regression (RR), random forest (RF), support vector machine (SVM) and back propagation (BP) were used to construct a cotton LAI prediction model.

The LAI of cotton exhibited an increasing and then decreasing pattern during the growth stage. Notably, the mean LAI values of cotton at the inner side of the budding stage, initial flowering stage, and boll setting stage were significantly greater than those at the lateral side (P<0.05). The selected indices exhibited significant correlations with each other across the periods (P<0.05). In general, the correlation between multi-spectral index and color index showed a decreasing trend as the growth stage progressed, and the selected indices were significantly correlated with cotton LAI in all stages (P<0.05), the correlation coefficients of multi-spectral index ranged from 0.35 to 0.85, and the correlation coefficients of color index ranged from 0.49 to 0.71, and those with a larger absolute value of the correlation coefficients were mostly multi-spectral indices, while those of the correlation coefficients of color index and cotton LAI were smaller. The estimated model performance results showed that the multi-spectral index was better than the color index in the cotton growth models, the predictive performance of the index models showed certain regularity with the change of growth, and NDVI was the optimal index for predicting cotton LAI. From the model results, the RF model and BP model obtained higher estimation accuracy under each growth stage. The LAI inversion model at the initial flowering stage had the highest accuracy, with the optimal model validation set R² of 0.809, MAE of 0.288, and NRMSE of 0.120. The optimal model validation set for the multiple growth stages had the R² of 0.386, MAE of 0.700, and NRMSE of 0.198.

There are significant differences in LAI between the inner and lateral sides of cotton during the budding stage, initial flowering stage, and boll setting stage. NDVI emerged as the optimal index for predicting cotton LAI at all growth stages, with the RF and BP models demonstrating superior performance. The effectiveness of the multiple growth stages model was notably lower compared to that of the single-growth model, with the optimal index identified as GNDVI and the optimal model as BP. The initial flowering stage appeared to be the optimal window for predicting cotton LAI. These findings can provide theoretical basis and technical support for utilizing UAV remote sensing to monitor cotton LAI.