|
The Satellite Application Facility on Land Surface Analysis (LSA SAF) produces and provides access to remotely sensed variables for the characterization of terrestrial ecosystems, such as land surface fluxes and biophysical parameters, taking full advantage of the EUMETSAT satellites and sensors. In this work, a procedure for the joint estimation of LSA SAF vegetation parameters is proposed. The approach includes the use of multi-task learning with gaussian processes (MTGP). The MTGP learns a shared covariance function on input features and a covariance matrix over tasks. Unlike the single output approaches, the proposed multi-task captures the inter-task dependencies among outputs. Two comparison exercises were undertaken to assess the estimation power of the MTGP as compared to single output algorithms such as standard gaussian processes regression (GPR), neural networks (NN), and random forest (RF). First, we evaluate the performance of MTGP in the context of deriving CO2 fluxes such as the gross primary production (GPP), net ecosystem exchange (NEE), and total ecosystem respiration (TER) blending SEVIRI/MSG and eddy covariance (EC) data. In addition, the MTGP prediction power was also assessed for the joint estimation of LAI, FAPAR, and FVC in a hybrid approach using radiative transfer model simulations and AVHRR/MetOp observations. The results show that MTGP outperforms the single output approaches in terms of accuracy. The MTGP multi-task optimization links outputs in such a way that the relationships among the biophysical parameters are better described obtaining a more robust model and therefore improving the accuracy of the estimates. The findings pave the way for future multi-task implementations in order to derive consistent outputs and accurate estimates of vegetation properties from remote sensing.
|
