Konferans bildirisi Açık Erişim

Digital soil mapping of AWC in arable lands: Comparison of different machine learning models

2021    Kaya, Fuat; Başayiğit, Levent; Dedeoğlu, Mert; Keshavarzi, Ali

The available water content (AWC) for plants can be defined as the amount of water retained in the soil resulting from the difference between the field capacity (FC) and the permanent wilting point (PWP). Both soil texture and structure affect the soil matrix potential on which the available water capacity to absorb water by plants depends. In arid and semi-arid environments, AWC has large variations in soil due to changing water balances and ongoing changes between water incomes and demand for vegetation type. Spatial highaccuracy estimation of available water capacity in water-scarce-affected areas such as the Mediterranean belt can support the efficient use of water. Digital Soil Mapping (DSM) is a method for the generation of spatial soil information through numerical models which are extracted spatial variations of soil properties from observation and environmental covariates that digitally represent soil formation factors. This study was designed to generate prediction maps for AWC using 3 different machine learning techniques: (i) Multiple Linear Regression (ii) Support Vector Regression (SVR), and (iii) Random Forest. Data collected from the Alluvial plain in Southwest Turkey (91 observations) were used to develop the models. In the estimating and map drawing of AWC, the spectral indices produced from Sentinel 2A images, topographical variables obtained from DEM, and the most recently CORINE land cover classes map were used as input parameters of the models. In the determination of mapping performance of the machine learning techniques for AWC, Lin's concordance correlation coefficient (LCCC) and root mean square error (RMSE) was used for data splitting (70%-30%) and the k-fold crossvalidation (n:5, repeated:3) was used. The results of the 3 models showed that the SVR model with higher LCCC (0.19) and lower RMSE (Test set: 2.72; crossvalidation: 2.33; %) was the best for AWC prediction. This study revealed that it can be applied by considering the comparative results of machine learning models to quickly predict and mapping of AWC using open-source accessible remotely sensed data. These digital maps can be used practically for monitoring soil water content, prioritizing irrigation schedules, and applied to other areas with similar environmental conditions.