Machine learning performance for predicting soil salinity using different combinations of geomorphometric covariates

- KED mostly outperformed ML and RM in mapping salt affected soils.
- ML, specifically DT, performed on par with KED in one of the study areas.
- An increase in spatial resolution showed no significant increase in accuracy.

Conventional methods of monitoring salt accumulation in irrigation schemes require regular field visits to collect soil samples for laboratory analysis. Identifying areas prone to salt accumulation by means of geomorphometry (i.e. terrain analyses using digital elevation models (DEMs)) can potentially save time and costs. This study evaluated the extent to which DEM derivatives and machine learning (ML) algorithms (k-nearest neighbour, support vector machine, decision tree (DT) and random forest) can be used for predicting the location and extent of salt-affected areas within the Vaalharts and Breede River irrigation schemes of South Africa. In accordance with local management policies, salt-affected areas were defined as regions with soil electrical conductivity (EC) values > 4 dS/m. Two DEMs, namely the one-arch second Shuttle Radar Topography Mission (SRTM) DEM and a photogrammetrically-extracted digital surface model (DSM), were used for deriving the derivatives. Wetness indices as well as hydrological and morphometric terrain analysis techniques were used to generate predictive variables. For comparative purposes, the predictive variables were also used as input to regression modelling and kriging with external drift (KED). Thresholds were applied to the regression models and KED results to obtain a binary classification. EC values based on in situ soil samples were used for model development, classifier training and accuracy assessment.The results show that KED achieved the highest overall accuracy (OA) in Vaalharts (79.6%), whereas KED and ML (DT) showed the most promise in the Breede River (75%). The findings suggest that the use of elevation data and its derivatives as input to geostatistics and ML holds much potential for monitoring salt accumulation in irrigated areas, particularly for simulating sub-surface conditions. More work is needed to investigate the potential of using ML and DEM-derivatives, along with other geospatial datasets such as satellite imagery (that have been shown to be effective for monitoring surface conditions), for the operational modelling of salt accumulation in large irrigation schemes.