ESTIMATION OF 20M LEFA AREA INDEX FROM LANDSAT AND SENTINEL-2 DATA USING A DEEP LEARNING APPROACH

Abstract

Global LAI satellite products have been generated at coarse spatial resolutions (250m or coarser). However, many applications urgently require spatiotemporal continuous vegetation biophysical products at fine scales (10-30m), which are not available worldwide. The USGS/NASA Landsat and the European Union Copernicus program's Sentinel-2 are the leading initiatives in medium-resolution land imaging. However, the optical Landsat (30m) and Sentinel-2 (20m) data are typically hampered by cloud contamination. Although great efforts have been made to retrieve 30m vegetation biophysical parameters from Landsat data, however, these single-phase based algorithms are sensitive to noises in the input reflectance data, resulting in gaps or fluctuations in time-series LAI products. In this study, we utilized LSTM deep learning methods to develop a model that can estimate continuous time-series LAI based on combined Landsat/Sentinel-2 surface reflectance data. Representative pixel samples for LAI training labels were derived from the 250m GLASS LAI datasets, which exhibit spatiotemporal continuity. Aggregated Landsat/Sentinel-2 surface reflectance data served as the training features. The accuracy of the estimated 20m LAI is validated against extensive public field measurements, with R2 of 0.7, RMSE of 0.9 for LAI. We applied this deep learning model to the Greater Bay Area (GBA) of China by constructing 20m Landsat/Sentinel-2 surface reflectance Analysis Ready Data. Finally, we assessed the derived seamless 20m LAI/FAPAR datasets by comparing them with coarse-resolution LAI/FAPAR products. The generated LAI 20m products at GBA are spatiotemporal continuous and consistent with the coarse resolution GLASS products with finer spatial resolution.