Above Ground Biomass Estimation from Passive Microwaves Brightness Temperatures Using Neural Networks

Above ground biomass (AGB) maps were estimated directly from microwave brightness temperatures (TB) using a machine learning approach. The accuracy of AGB retrievals from Artificial Neural Networks (ANN) is explored using both a multi-angular (using TBs products from the SMOS mission) and multi-frequency approach (using multi frequency measurements from the AMSR-E mission), an additional ANN inversion including optical indexes (MODIS-NDVI) is also discussed. Higher incidence angles have been shown to provide more information during the inversion process for AGB estimates than lower angles. Retrievals from the multi-angular lower-frequency inversion (SMOS - 1.4GHz) performed better than any individual higher-frequency retrievals. The addition of multi-frequency TBs (AMSR-E) to lower-frequency multi-angular TBs improves the performance of ANN models (from
$\mathrm{R}^{2}\approx 0.94$
to 0.96). Adding MODIS-NDVI to the inversion process improves the performance in an additional ~0.05%.