A comparison of methods to estimate leaf area index using either crop-specific or generic proximal hyperspectral datasets

The leaf area index (LAI) is an important parameter indicating the crop growth status. Many vegetation indices and models have been developed to estimate LAI of different crops. However, the utility of and differences between crop-specific and generic algorithms for LAI estimation covering several crops, dates, and sites need to be compared. The main objectives of this study were to: (1) evaluate the suitability and robustness of different vegetation indices for LAI estimation of different crops; (2) compare the performance of crop-specific and generic algorithms; (3) evaluate the performance of statistical regression, machine learning regression, and the PROSAIL-D with different inversion strategies. Results showed that: (1) the simple ratio index (SR) performed best for cotton and winter wheat, the MERIS terrestrial chlorophyll index (MTCI) performed best for maize; (2) the models trained over specific crop types performed better than those trained over all crop types together; (3) for crop-specific models, artificial neural networks (ANN) performed better than support vector machine regression (SVR), partial least square regression (PLSR), and statistical regression methods, while the PROSAIL-D yielded similar or slightly better performance as compared with ANN method. Furthermore, the look-up table (LUT) strategy performed better than iterative optimization strategy (Shuffled Complex Evolution method developed at the University of Arizona, SCE-UA). Results indicated that there was not an optimal model that could be used to precisely estimate LAI of different crops since the relationship between the specific hyperspectral reflectance and the LAI of different crops were different. In addition, the PROSAIL-D with crop specific input parameters is a potential method for LAI estimation of different crops with hyperspectral reflectance.

Automated summary

This study evaluated the suitability and robustness of different vegetation indices for LAI estimation of different crops, compared the performance of crop-specific and generic algorithms, and evaluated the performance of statistical regression, machine learning regression, and the PROSAIL-D with different inversion strategies. The results showed that different crops require different vegetation indices and algorithms for accurate LAI estimation, and the PROSAIL-D with crop specific input parameters is a potential method for LAI estimation of different crops.