Predição da textura do solo com sensores proximais e atributos de terreno com diferentes resoluções espaciais

Abstract

Soil texture is a key factor influencing various aspects of the terrestrial environment, including fertility, water retention, and biodiversity. Understanding its distribution is therefore essential for sustainable agricultural land management. With the advent of proximal and remote sensing technologies, it has become possible to obtain detailed information about soil texture with greater efficiency. These technologies enable faster and more cost-effective soil analysis, optimizing analytical processes. In this context, the study initially evaluated the contribution of proximal and remote sensing data in predicting soil texture. To achieve this, 114 samples were collected from the surface and subsurface horizons of the Muquém Farm at UFLA. These samples were analyzed using the pipette method for granulometry, complemented by measurements from portable X-ray fluorescence spectrometry and magnetic susceptibility. Fifteen (15) terrain attributes were derived from Digital Terrain Models with a resolution of 12.5 m, using QGIS software. Prediction models were built using the Random Forest algorithm, demonstrating that integrating proximal sensor data with topographic information significantly improves the accuracy of soil property maps. Secondly, the study analyzed the impact of terrain attributes and parent material at different spatial resolutions on soil texture prediction at the UFLA campus. DTMs derived from contour lines at resolutions ranging from 1 m to 30 m, as well as data from the Alos Palsar and SRTM projects, were used to generate 15 terrain attributes. The models were trained using the Random Forest and SVM algorithms based on 232 soil samples. The analysis revealed that parent material was a determining factor in predicting sand, silt, and clay, followed by attributes such as Analytical Hillshading, LS-Factor, Channel Network Base Level, Multiresolution Index of Ridge Top Flatness, and Multiresolution Index of Valley Bottom Flatness. Additionally, DTMs with resolutions between 10 m and 20 m produced the most accurate models, confirming that higher resolutions (1m, 5m) do not necessarily guarantee better predictive performance. This study demonstrated that the combination of proximal sensing data, DTMs, and machine learning algorithms constitutes an effective approach for digital soil mapping, promoting sustainable management of soil resources.