Title:	Integrating GIS and remote sensing for soil attributes mapping in degraded pastures of the Brazilian Cerrado.
Authors:	LOUZADA, R. O. BERGIER, I. BOLFE, E. L. BARBEDO, J. G. A.
Affiliation:	RÔMULLO OLIVEIRA LOUZADA, INSTITUTO DE MEIO AMBIENTE DE MATO GROSSO DO SUL; IVAN BERGIER TAVARES DE LIMA, CNPTIA; EDSON LUIS BOLFE, CNPTIA; JAYME GARCIA ARNAL BARBEDO, CNPTIA.
Date Issued:	2025
Citation:	Soil Advances, v. 3, 100044, 2025.
Description:	Abstract: Remote sensing is a crucial tool for soil assessment, yet soil complexity and sensor limitations hinder accurate analysis. This study integrates active and passive remote sensing data with Machine Learning (ML) methods to predict the physicochemical properties of degraded sandy soils in the Brazilian Cerrado. The 1197 ha area was divided into management zones. Soil samples were collected from each management zone at 0–0.2 m and 0.2–0.4 m depths. The samples were then bulked (n = 99) and analyzed for texture (clay, silt, sand), pH, soil organic matter (SOM), cation exchange capacity (CEC), effective CEC (ECEC), base saturation (V), and macro- and micronutrients (e.g., Ca, Mg, K, Fe, Mn). Composite samples from management zones, were matched with 128 orbital variables from Sentinel-1, Sentinel-2 (2023), and ALOS-PALSAR-1. The variables include spectral bands, vegetation and soil indices, gray-level co-occurrence matrices (GLCM), backscatter coefficients, polarimetric decompositions, and topographic indices. A key innovation was evaluating statistical metrics beyond the mean—such as medians, sums, and variances—within MZs. The models were processed using Random Forest (RF), with variable selection assessed via the Boruta algorithm. The tested approaches included (T1) RF with mean-based variables, (T2) RF + Boruta, (T3) RF with the highest correlation metrics, and (T4) RF + Boruta with correlation-based metrics. Results showed that Boruta-enhanced models (T2 + T4) improved performance in 89 % of cases. Correlation-based metrics (T3/T4) were more effective in 72 % of models than mean-based approaches (T1/T2). The best models demonstrated high accuracy for clay (R² = 0.81; RMSE = 25.2 %), CEC (0.73; 23.6 %), silt (0.71; 44.7 %), and K (0.62; 44.3 %) in the 0–0.2 m layer. In the 0.2–0.4 m layer, top-performing attributes included clay (R² = 0.86; RMSE = 19.1 %), sand (0.78; 10.6 %), silt (0.76; 39.3 %), and SOM (0.68; 21 %). Elevation and GLCM metrics emerged as key predictors across depths. These findings highlight the effectiveness of integrating diverse remote sensing data with ML for soil attributes mapping, particularly for clay and CEC.
Thesagro:	Sensoriamento Remoto Sistema de Informação Geográfica Análise do Solo Textura do Solo Matéria Orgânica
NAL Thesaurus:	Remote sensing Geographic information systems Soil analysis Soil texture Subsurface soil layers
Keywords:	Random Forest Statistical summarization Machine Learning
ISSN:	2950-2896
DOI:	https://doi.org/10.1016/j.soilad.2025.100044
Type of Material:	Artigo de periódico
Access:	openAccess
Appears in Collections:	Artigo em periódico indexado (CNPTIA)