The use of visible-near infrared spectroscopy (VNIRS) and machine learning techniques for the prediction of some physical and chemical properties of Upper Tigris Basin soils

Spatial change maps are valuable tools for the accurate management of agricultural lands and optimal use of inputs, as they readily identify locations with the lowest and highest levels of parameters such as clay, sand, silt, EC, lime, organic matter, Zn, Cu, Mn, and Fe elements, which have significant effects on the quality and yield of crops in agricultural production. However, obtaining these analyses poses challenges such as time, cost, labor, etc. To address these issues, in addition to mapping, faster results can be obtained through Near-Infrared Spectroscopy (NIRS) techniques (ranging from 350nm to 2500nm). The aim of this study is to compare the traditional soil analyses with Visible and Near-Infrared Spectroscopy (VNIRS) analysis (350nm-2500nm) for some analyses of soils located in the Upper Tigris Basin using machine learning. For this purpose, soil samples were collected from 386 different points within the study area, covering approximately 1,000,000 hectares in the Upper Tigris Basin, at depths of 0-20 cm, and scanned using VNIRS. After VNIRS scanning, spectral readings were averaged over 10-25-50 nm. The best average rate for each soil property was determined by applying Support Vector Regression, Random Forest Regression, Decision Tree Regression, and XGBoost Regression models. The best model among these combinations was identified. It was observed that the averaged spectra , specifically spectra averaged over 50-nm, produced the best predictions for clay, silt, Fe, Cu, and Mn contents. Additionally, the XGBoost Regression model emerged as the best-performing model. Furthermore, examining the model accuracy indicators R2, RMSE, and RPD, it is determined that, for clay content, the R2 value is 0.93, RMSE value is 0.02, and RPD value is 3.92. For silt content, the R2 value is 0.95, RMSE value is 0.02, and RPD value is 4.32. Similarly, for Fe content, the R2 value is 0.95, RMSE value is 0.02, and RPD value is 4.49. For Cu content, the R2 value is 0.95, RMSE value is 0.02, and RPD value is 4.52. Lastly, for Mn content, the R2 value is 0.95, RMSE value is 0.02, and RPD value is 4.64. When examining sand, EC, organic matter, lime, and Zn parameters, it was observed that the spectra averaged over 10 nm produced the best predictions. Additionally, the XGBoost Regression model was the best-performing model for these parameters. When examining the model accuracy for other variables, it is determined that, for sand content, the R2 value is 0.94, RMSE value is 0.02, and RPD value is 4.06. For EC content, the R2 value is 0.94, RMSE value is 0.02, and RPD value is 4.13. Similarly, for CaCO3 content, the R2 value is 0.94, RMSE value is 0.02, and RPD value is 3.95. Regarding organic matter content, the R2 value is 0.92, RMSE value is 0.03, and RPD value is 3.56. For Zn content, the R2 value is 0.95, RMSE value is 0.02, and RPD value is 4.51. Furthermore, as the analyses for clay, sand, silt, EC, CaCO3, and organic matter had already been established under the TUBITAK project, this thesis concentrated on determining the missing micro-nutrient element contents and creating maps depicting distance-dependent variations.