Mapping clay fraction oxides in Brazil using Earth observation strategy

The major oxides in the clay fraction of tropical soils are iron (Fe2O3), aluminum (Al2O3), and silicon (SiO2) oxides, which are responsible for the soil’s capacity to provide multiple ecosystem services. Therefore, they are used to classify the soils into different pedological classes. Despite their importance of these oxides, quantifying them on a large scale presents significant challenges. The most common method is laboratory sulfuric acid digestion, which is expensive, complex, and environmentally detrimental. To overcome these issues and provide faster results, we developed a satellite technique associated with machine learning (ML) to map Fe₂O₃, Al₂O₃, and SiO₂ in all agricultural areas in Brazil at 30 m resolution. Additionally, we tested the feasibility of the generated maps to infer soil weathering, and assist in the construction of pedological maps. A dataset, comprising 5,330 sites (0–20 cm and 80–100 cm) across all 27 states was employed in prediction. Six spectral variables obtained from the historical Landsat series (bare soil) and seven terrain attributes derived from a digital elevation model were employed to generate the Fe₂O₃, Al₂O₃, and SiO₂ maps, using the Random Forest algorithm. The predicted maps of oxides covered nearly 3.48 million km² (∼40 % of the national territory). The best predictions were observed for Fe₂O₃ in the 0–20 cm layer (RMSE = 49.8 g.kg⁻¹, RPIQ = 1.82, and R² = 0.62), while the worst predictions were for SiO₂ in the 80–100 cm layer (RMSE = 65.3 g.kg⁻¹, RPIQ = 1.50 and R² = 0.22). It was possible to infer soil weathering using the Ki index. Despite the models not showing such high R² values, the results are aligned with legacy maps, highly weathered soils were observed in the plateaus of the Cerrado biome, while younger soils were observed in the arid Caatinga biome and waterlogged soils in the Pantanal biome. The generated maps also demonstrated a high potential for grouping pedological soil classes. They also revealed a relationship between oxide contents and the NDVI of sugarcane crops, indicating potential applications in crop management. Moreover, this satellite-based technique, supported by ML, presents a plausible approach to predict oxide fraction at high spatial resolution for large areas.