Combining both spectral and textural indices for alleviating saturation problem in forest LAI estimation using Sentinel-2 data

Abstract

Leaf area index (LAI) is an important structural parameter that controls energy, water, and gas exchanges of plant ecosystems. At present, remote sensing techniques, especially passive optical remote sensing, are the most widely used approaches to estimate LAI. However, the prevalent usage of spectral information alone often saturates, which is a major issue for retrieving LAI in particular for forests. In this study, our goal was to improve LAI estimation using Sentinel-2 data by combining both spectral and texture information to minimize the saturation effect in a typical cool-temperate deciduous forest in Japan. Two LAI field datasets obtained using digital hemispherical photographs, representing a range of LAI conditions, were used to investigate the performance of spectral and texture features or a combination of both. Texture information was derived using the grey-level co-occurrence matrix (GLCM) method. Commonly used vegetation indices and newly developed textural indices were computed for comparison. Spatial autocorrelation was determined prior to exploratory regression analyses to explore methods for estimating LAI using a combination of spectral- and texture-based features. Regression methods included generalized least squares (GLS), stepwise, principal components (PCR), and partial least squares (PLS). The results demonstrate that the texture information derived from Sentinel-2 imagery helps estimate LAI by minimizing the saturation problem, especially in the period of maximum LAI values. The combination of both features using the PLS model provided a higher accuracy estimation scheme for LAI (R 2: 0.66–0.75, RMSE: 0.23–0.37) and we foresee the wide application of the approach especially for a closed-canopy forest facing serious saturation problem when deriving LAI from reflected spectral information.