Publications

Zhang, GD; Yin, GF; Zhao, W; Wang, ML; Verger, A (2025). A deep learning method for generating gap-free FAPAR time series from Landsat data. REMOTE SENSING OF ENVIRONMENT, 326, 114783.

Abstract
Fraction of Absorbed Photosynthetically Active Radiation (FAPAR) is a key indicator of photosynthetic activity and primary productivity in terrestrial ecosystems. While moderate-coarse spatial resolution FAPAR products have enabled global vegetation studies, their pixel sizes smooth fine-scale heterogeneity and limit applications needing a detailed spatial characterization. Landsat provides multispectral data at 30 m spatial resolution enabling global and long-term FAPAR estimation at finer spatial detail. But challenges exist in deriving continuous Landsat FAPAR time series due to infrequent clear observations. We propose here a generic two-step method to produce gap-free Landsat FAPAR time series. First, in clear-sky conditions, FAPAR is retrieved from Landsat surface reflectance observations using a random forest (RF) regression model previously trained with the Global land surface satellite (GLASS) V6 FAPAR product. In a second step, a novel Bidirectional Temporal Convolutional Network with Sparrow Search Algorithm and Attention mechanism (SSA-BiTCN-Attention) was used to reconstruct the missing FAPAR values. The temporal information of Landsat clear-sky FAPAR within a five-year window was used for predicting the missing values. The reconstruction model accurately predicted missing FAPAR values across different land cover types with RMSE ranging from 0.08 to 0.12. The validation results showed a good agreement between the estimated Landsat FAPAR and ground measurements from VAlidation of Land European Remote Sensing Instruments (VALERI), ImagineS and Ground Based Observations for Validation (GBOV), with R2 values ranging from 0.82 to 0.92, RMSE values from 0.10 to 0.12, and bias values from 0.02 to 0.05. The Landsat retrievals are consistent with GLASS and Moderate-Resolution Imaging Spectroradiometer (MODIS) FAPAR products and improve these two products in terms of accuracy and spatial resolution. As a demonstration case study, we applied our method to generate 30-m, 16-day FAPAR time series from 2013 to 2023 over China. The Landsat gap-free FAPAR product exhibits seamless spatial coverage and temporal continuity across China. This innovative method has the potential to be applied to multiple satellite data and land surface products to generate gap-free high spatio-temporal time series of land surface variables at the global scale, which will contribute in improving environmental modeling, carbon cycling studies, and agricultural applications.

DOI:
10.1016/j.rse.2025.114783

ISSN:
1879-0704