Mapeamento da biomassa acima do solo na bacia do Rio Grande utilizando dados de sensoriamento remoto e aprendizado de máquina

Abstract

Aboveground biomass (AGB) is one of the most dynamic carbon reservoirs in the terrestrial ecosystem, but little is known about its spatial distribution. Currently, the most accurate methods for quantifying AGB are expensive, time-consuming, and limited to small areas. In this scenario, remote sensing (RS) appears as a promising tool, providing sufficient spatial and temporal coverage for continuous and low-cost monitoring of AGB in large areas. The objective of this work was to map the AGB in the Rio Grande basin, Minas Gerais, from multi-sensor RS data, such as information from the B2-5 (surface reflectance) and B10 (surface temperature) bands of the sensors OLI and TIRS from the Landsat 8 (L8) satellite, Synthetic Aperture Radar (SAR) data from Sentinel-1A (S1), and data from the SRTM digital elevation model. We also used data derived from these, such as five vegetation indices (NDVI, SAVI, DVI, ARVI, and EVI) and 18 texture measures derived from Gray-level Co-occurrence Matrix (GLCM), using a 3 x 3 window size. To perform the modeling and mapping of the AGB, we used the Random Forest machine learning algorithm. We used the nested cross-validation method and the Root Mean Square Error (RMSE) metric to measure the performance of the model, which presented a value of 48.79 Mg.ha-1 (44.22%). We optimized the mtry (22), sample.fraction (0.22) and min.node.size (10) hyperparameters using the random search method. The results reinforced the importance of using data from different sensors and domains (spectral and spatial) in AGB modeling. The ARVI vegetation index wasthe most important variable for the model. However, the variable could not capture the change in AGB in areas of denser vegetation, with AGB values close to 114 Mg.ha-1

. The variables elevation and the texture metrics dissimilarity and sum entropy associated with VH polarization (S1) and DVI vegetation index (L8) were essential to mitigate the effect of data saturation.