Determining the accuracy of the landsat-based land continuous Variable Estimator

Abstract

The inversion framework Land continuous Variable Estimator for Landsat data (LoVE)-Landsat has been recently developed, but the estimation accuracy has not been well determined. LoVE-Landsat is a data-assimilation-based inversion framework capable of estimating a series of daily 30-m spatiotemporal continuous land surface variables from Landsat top-of-atmosphere (TOA) data. This paper presents the comprehensive validation results of a land surface variables estimation conducted through LoVE-Landsat, including the leaf area index (LAI), fraction of absorbed photosynthetically active radiation (FAPAR), surface broadband albedo, incident photosynthetically active radiation (PAR), and incident shortwave radiation (ISR). This validation involved a direct validation using extensive ground measurements and an inter-comparison with existing satellite products. In situ measurements came from a total of 196 sites covering different biome types from seven networks distributed around the world. Among these sites, more than 2000 LAI and FAPAR plot measurements representing the Landsat pixel size were collected from 40 sites of the Bigfoot, VALERI, and ImagineS networks, about 100 LAI and FAPAR reference values at 3-km scale were provided by 52 sites of DIRECT datasets, and nearly 70,000 albedo and ISR and 40,000 PAR tower-based measurements were gathered from 143 sites of Ameriflux, Euroflux, Ozflux, and SURFRAD networks. Results showed that the LoVE-Landsat 30 m LAI estimates were accurate, with R2 = 0.76 and root-mean-square-error (RMSE) = 0.89, and were slightly more accurate than the coarse resolution Global LAnd Surface Satellite (GLASS) LAI product at the kilometer scale (RMSE = 0.75 vs. 0.79). Validation at the 143 flux network sites showed that the daily snow-free LoVE-Landsat 30 m shortwave albedo had an accuracy comparable to the MODIS daily albedo product (RMSE around 0.037) and higher accuracy (RMSE = 0.031) when Landsat had clear-sky observations. Direct validation of ISR and PAR estimation showed high accuracy, with RMSE values of 102.8 and 48.7 W/m2, respectively. Spatial and temporal evaluation at six typical sites also showed that LoVE-Landsat could produce consistent spatially and temporally continuous estimation. Although a more comprehensive validation of all retrievals still needs to be conducted, both the direct validation and product comparison results of the key variables indicate that LoVE-Landsat can estimate a group of spatially and temporally continuous variables from Landsat observations accurately, which demonstrates the strong potential of LoVE-Landsat to generate global 30-m continuous land products.