Comparative analysis of LAI retrieval from hyperspectral imagery using machine learning approaches

< p >Vegetation biophysical and biochemical variables are key inputs to a wide range of modelling approaches for carbon, water, energy cycle, climate and agricultural applications. Leaf Area Index (LAI) is among the most important canopy variables, used by many different physiological and functional plant models. Several approaches have been developed for vegetation properties retrieval from remotely sensed hyperspectral data. Among them, nonparametric machine learning methods have increasingly gained attention in vegetation variable retrieval due to their flexibility and efficiency while working with data of high dimensionality over the last decades. Although these methods provide reasonable accuracy at relatively high speed, they are mainly restricted to estimate values within their training domain and often perform poorly on the marginal values (i.e. outside of the training domain). The performance of these methods has not been adequately studied in retrieving LAI on the marginal values. This study employs four well-known machine learning methods including SVR, GPR, ANN, and RF to retrieve LAI from a hyperspectral CHRIS-Proba image over Barrax, Spain, in order to inspect their capability in retrieving marginal values. The results showed that although all the methods perform similarly well on retrieving LAI over the training domain values with RMSE values of less than 0.5 and relative error of less than 10%, GPR and SVR performed slightly better. However, ANN outperformed the other methods in estimating LAI on the marginal values, resulted in the generated LAI map more consistent with the NDVI map, as well as, the hyperspectral image of the region.