Prediction of soil properties via portable x-ray fluorescence (pXRF) spectrometry in Brazil

• Autor(a) principal: Pelegrino, Marcelo Henrique Procópio
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/33940
• topic: Aprendizagem de máquina; Espectrometria de fluorescência de raios-X portátil (pXRF); Mapeamento digital do solo; Machine learning; Digital soil mapping; Ciência do Solo

Soils are the main substrate for food production. Increasing environmental demand and pressure imposes greater productivity, profitability, and mitigation of environmental impacts in the stages and production techniques. From this perspective, the importance of knowing the chemical, physical, and biological soil properties is evident. For agriculture, a better understanding of soil fertility enables a more rational use of resources and inputs in crop planning. However, the acquisition of this knowledge requires soil sampling and analyses, which increases in number and volume as knowledge is refined. This makes the process expensive.. Moreover, the usual point-sampling approach limits understanding of the spatial and temporal variability of soil properties. In this sense, the acquisition of data through remote (e.g., satellite images) and proximal (e.g., portable spectrometers) sensors has refined and complemented the knowledge about soil properties with the aid of computational modeling. As a widely diffused source in environmental modeling, one can easily obtain terrain attributes (e.g., topographic wetness index) from digital elevation models (DEM) in GIS environments. Regarding proximal sensors, the portable X-ray fluorescence (pXRF) spectrometry has the advantages of ease, speed, and non-generation of waste in its operation, as well as the advantages of being portable. The present dissertation is divided in two chapters, whose objectives are: modeling and spatial prediction of the available micronutrients contents Fe, Mn, Cu, and Zn, through data obtained from terrain attributes (TA), pXRF, and parent material information (PM), for surface and subsurface horizons separately and combined (n = 153), in different combinations of datasets and spatial resolution, using the random forest (RF) algorithm; and modeling and spatial prediction of the available levels of the macronutrients P, Ca and K, through the pXRF sensor data for the surface horizon (n= 90), using simple linear regression (LR), polynomial regression (PR), power regression (PwR), multiple linear regression (SMLR) and random forest (RF). The study area is located between longitudes 501031 and 504192 mE and latitudes 7651139 and 7653537 mN, zone 23 K, located on the campus of the Federal University of Lavras, with approximately 315 ha. Its soils are developed from gneiss, gabbro and alluvial sediments. The climate is Cwa according to Köppen classification system, with average annual temperature of 20.4 °C and average annual rainfall of 1.460 mm. Samples were collected on a regular grid design of 200 m between sampling places., Samples were submitted to laboratory analysis to determine the respective nutrients. Subsequently, a portion of each sample was analyzed on a pXRF model S1 Titan LE (Bruker Nano Analytics, Kennewick, WA, USA) in Trace mode for 60 s in triplicate. The TA were generated with the SAGA GIS software from 5 and 10 m resolution DEM. The data were separated into training (70%) and validation sets (30%), and the models were generated in R software (RF) and SigmaPlot (LR, PR, PwR and SMLR). For the purposes of analysis and comparison between the models, we used the coefficient of determination (R2), adjusted R2 (R2adj), root mean squared error (RMSE), normalized root mean squared error (nRMSE) and mean error (ME) for the first chapter, and R2, RMSE, mean absolute error (MAE) and the residual deviation of predictions (RPD) for the second chapter. After determination of the best models, the spatial prediction was followed to generate the available nutrient map. The variables of the pXRF when present in the model were spatialized for the entire area through inverse distance weighting (IDW) interpolation. The 10 m TA were better than the 5 m resolution for predictions. It was possible to obtain good results in the spatial prediction of available Fe using only 10 m TA (R2 = 0.88; RMSE = 59.97 mg kg-1 and ME = 24.00 mg kg-1) and for the others with pXRF + 10 m TA + PM (0.85; 29.65 mg kg-1; 9.70 mg kg-1 for Mn, 0.64; 3.11 mg kg-1; 0.71 mg kg-1 for Zn e 0.82; 1.17 mg kg-1; 0.43 mg kg-1 for Cu, respectively) In the predictions of the macronutrients, the PwR approach obtained the best results (R2 = 0.80 and RMSE = 1.63 cmolc dm-3 for exchangeable Ca2+, and 0.53 and 6.92 mg dm-3 for available P). It was not possible to establish a correlation between the available K+ contents and the total K2O content provided by pXRF. Proximal sensor data associated with TA data can accurately predict exchangeable/available nutrient contents in tropical soils.