Work

The Brazilian intelligent cloud soil analysis system for monitoring, reporting, and solution verification

Year: 2026

Type: article

Abstract: As climate-smart agriculture and soil monitoring frameworks expand globally, scalable analytical services are increasingly needed. Building on the Brazilian Soil Spectral Library (BSSL) and the original Brazilian Soil Spectral Service (BraSpecS) platform, we developed BraSpecS 2.0, an automated cloud-based platform leveraging artificial intelligence to deliver near-real-time soil estimates directly from Vis–NIR–SWIR and MIR spectral measurements. The system advances beyond operator-dependent processing by implementing an automated Long Short-Term Memory (LSTM) autoencoder that learns a compact latent spectral representation from the BSSL. Property-specific XGBoost regression models then use these latent embeddings for multi-property prediction. The evaluation focused on primary spectrally active soil attributes, including soil organic carbon (SOC), texture fractions (clay, silt, and sand), and iron and titanium oxides. Among these targets, clay and Fe 2 O 3 in Vis–NIR–SWIR and SOC in MIR achieved R 2 ≥ 0.80, with strong predictive performance. Conversely, available P and exchangeable K achieved low prediction performance in both spectral domains (R 2 ≤ 0.44), indicating no practical benefit of infrared spectroscopy for the estimation of these attributes; they are reported only to delimit the boundaries of the approach. External validation across independent datasets from seven laboratories demonstrated the potential for model generalization of clay and SOC while revealing inter-laboratory variability and emphasizing the need for cross-instrument harmonization. BraSpecS 2.0 establishes an automated and scalable platform for national soil monitoring initiatives.

Source: Geoderma

DOI: 10.1016/j.geoderma.2026.118100