Examinando por Autor "Manda, Godwell"

Mostrando 1 - 1 de 1

Restringido
A New approach to predict nutrient content in soil using NIR spectroscopy and machine learning.
(Universidad EARTH, 2020-12) González Jiménez, Tatiana Magali; Manda, Godwell; Perret, Johan
The visible and near-infrared reflectance spectroscopy (VNIRS) combined with machine learning can be used to predict soil nutrients. This research was undertaken to test and improve chemometric methods to produce the most accurate soil infrared spectral models to predict the nutrient content. A FieldSpec 4 Standar-Res ASD (Analytical Spectral Devices Inc., Boulder, CO, USA) spectroradiometer with a spectral range of 350 nm to 2500 nm with a spectral resolution of 3 nm and 10 nm in the ranges of 350 nm to 1000 nm and 1000 nm and 2500 nm respectively was used to generate the Vis-NIR spectral signatures. A total of 2318 soil samples since 2017 were collected, processed (oven-dried and grind) where macronutrients (P, K, S, Ca, Mg) and micronutrients (Zn, Cu, Fe, Mn, B), Si, and Na were analyzed in the laboratory. Further, Carbon, Nitrogen, soil pH, soil extractable acidity (EA), soil texture, bulk density, soil effective Exchange Capacity, and Soil organic content were assessed. The same samples were scanned with a spectroradiometer. Then, the spectral data were preprocessed using different techniques. Four machine learning methods namely, Memory based learning (MBL), Support Vector Machine (SVM), (RF), and Partial Least Square Regression (PLSR) were employed to construct soil nutrient, prediction models. Results indicated that the MBL model outperformed the RF, SVM, and PLSR. The 28 best models of all the property are measured with R² between 0.72-0.94. From this, 82 % (23 best models) were dominated by MBL while 18 % (5 best models) RF, SVM, and PLSR. The highest prediction result was obtained by MBL on the element Fe with (R²) = 0.94, a ratio of performance to deviation (RPD) = 4.08, and root mean square error of prediction (RMSEP) = 14.83 %). It can be concluded that Machine-Based Learning can be used to predict the nutrients of Costa Rican soils.