L4F - Training dataset and layers for biomass prediction

Those data represent the Brazilian Amazon by 68,629,072 250-m pixels, and 141,032 pixels have LiDAR information and therefore had the AGB estimated, and converted to Mg ha-1 (file trainning_dataset.csv). To generate a wall-to-wall map of the Brazilian Amazon at 250-m resolution, we trained a Random Forest (RF) model 29 using the AGB estimated pixels and remote sensing layers formed by: MODIS vegetation indices, Shuttle Radar Topography Mission (SRTM) data, Tropical Rainfall Measuring Mission (TRMM), and Phased Array type L-band Synthetic Aperture Radar (PALSAR) data, along with the central coordinates of each 250m pixel, stored in amazon_*.csv files.

Derived from MODIS, we used the Vegetation Indices 16-Day L3 Global 250m temporal series (MOD13Q1) products from 2016, including the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI), from MODIS. From TRMM we used the 3B43 V6 precipitation data, with each pixel value representing the monthly accumulated precipitation from 1998 to 2016 at a resolution of 0.25 degrees. From PALSAR, we used the L-band image in the HH and HV polarizations, acquired in 2015. When necessary, the remote sensing products were resampled by means to a 250 m grid.

Random Forest models were tested using the H2O_Flow platform and produced the best model based on RMSE and R², containing NDVI q3, PALSAR HV, TRMM mean, X, SRTM, Y, PALSAR HH, EVI q1, EVI mean, NDVI mean, NDVI q1, EVI q3.