Decameter Cropland LAI/FPAR Estimation From Sentinel-2 Imagery Using Google Earth Engine

Abstract

Leaf area index (LAI) and fraction of photosynthetically active radiation (FPAR) products at regional and global scales have already been extensively and routinely generated from medium-resolution sensors. However, there is a lack of high-resolution LAI/FPAR product, which is especially essential for crop growth and drought monitoring of cropland in patches. This article proposes a processing framework for the derivation of decameter cropland LAI and FPAR in the Northern China plain from Sentinel-2 surface reflectance data with a random forest (RF) algorithm by exploiting the capabilities of the Google Earth Engine (GEE) cloud platform. The training database is generated from the spatially aggregated Sentinel-2 surface reflectance and the corresponding Moderate Resolution Imaging Spectroradiometer (MODIS) LAI/FPAR product over homogeneous cropland, and the training samples are strictly filtered for the best quality. RF is then trained over the processed Sentinel-2 surface reflectance and the filtered MODIS LAI/FPAR under two input groups—one group is for Sentinel-2 spectral bands of 10-m resolution only, and the other group supplements the Sentinel-2 red-edge (RE) and shortwave infrared (SWIR) bands of 20-m resolution. Extensive comparisons and validation are carried out, and they demonstrate that the new method can generate spatial and temporal consistent LAI/FPAR with MODIS at high spatial resolution. The retrieval accuracy is slightly better for 20-m input groups than that for 10-m input groups, confirming the value of RE and/or SWIR in cropland LAI/FPAR estimate. This article also demonstrates that GEE is a suitable high-performance processing tool for high-resolution biophysical variables estimation.