Publicações

Due to the limited coverage of field Aboveground Biomass (AGB), remote sensing becomes an alternative for monitoring carbon stocks at the landscape scale. However, the most commonly used sensors have limited spectral resolution. Hyperspectral imaging (HSI) provides high-resolution information, although its high data dimensionality becomes a challenge for modeling. In this context, selection of suitable variables is a critical step for estimating AGB from HSI data. Support Vector Regression coupled with the Recursive Feature Elimination approach (SVR-RFE) can produce parsimonious models from a reduced subset of features. We applied the SVR-RFE in a 5-fold cross-validation strategy with 5 repetitions to determine which hyperspectral variables were most effective to estimate AGB. We used field AGB from 147 inventory plots across the Brazilian Amazon and 64 plot-level HSI metrics, including 14 reflectance bands, 30 vegetation indices, continuum-removal absorption features at five wavelengths (495, 670, 980, 1200, and 2100 nm), and endmember fractions (green vegetation, shade, and non-photosynthetic vegetation/soil) from Spectral Mixture Analysis. The SVR-RFE explained 67% of the AGB variation, by selecting eight HSI variables. The three most effective variables came from the shortwave infrared region (width and depth of the 2100-nm absorption band and the NDNI index), related to canopy moisture and lignin-cellulose-nitrogen absorption bands. Four metrics were retrieved from the water absorption band centered at 980 nm (depth, asymmetry, and the indices PWI and LWVI1). The width of the band placed at 495 nm was also selected. SVR-RFE proved to be an efficient technique for estimating AGB from HSI data.