Spatial Modeling of a Soil Fertility Index using Visible–Near-Infrared Spectra and Terrain Attributes

Our objective was to develop a methodology to predict soil fertility using visible–near-infrared (vis–NIR) diffuse reflectance spectra and terrain attributes derived from a digital elevation model (DEM). Specifically, our aims were to: (i) assemble a minimum data set to develop a soil fertility index for sugarcane (Saccharum officinarum L.) (SFI-SC) for biofuel production in tropical soils; (ii) construct a model to predict the SFI-SC using soil vis–NIR spectra and terrain attributes; and (iii) produce a soil fertility map for our study area and assess it by comparing it with a green vegetation index (GVI). The study area was 185 ha located in São Paulo State, Brazil. In total, 184 soil samples were collected and analyzed for a range of soil chemical and physical properties. Their vis–NIR spectra were collected from 400 to 2500 nm. The Shuttle Radar Topographic Mission 3-arcsec (90-m resolution) DEM of the area was used to derive 17 terrain attributes. A minimum data set of soil properties was selected to develop the SFI-SC. The SFI-SC consisted of three classes: Class 1, the highly fertile soils; Class 2, the fertile soils; and Class 3, the least fertile soils. It was derived heuristically with conditionals and using expert knowledge. The index was modeled with the spectra and terrain data using cross-validated decision trees. The cross-validation of the model correctly predicted Class 1 in 75% of cases, Class 2 in 61%, and Class 3 in 65%. A fertility map was derived for the study area and compared with a map of the GVI. Our approach offers a methodology that incorporates expert knowledge to derive the SFI-SC and uses a versatile spectro-spatial methodology that may be implemented for rapid and accurate determination of soil fertility and better exploration of areas suitable for production.