Application of artificial neural networks and partial least squares regression to predict irrigated land soil salinity in the Rift Valley Region, Ethiopia

Study region: Rift Valley-Awash River Basin, EthiopiaStudy focus: Irrigation schemes in Awash Basin, Ethiopia, are severely affected by the buildup of soil salinity. The main source of the salinity is shallow groundwater heightened by improper irrigation practices and management. However, salinity predictions have not been developed based on direct measured data for the basin. Therefore, this study aims predicting topsoil salinity in irrigated land from basic hydrological parameters using two approaches: Artificial neural networks (ANNs) and Partial least squares regression (PLSR). Irrigation water amount, water table depth, precipitation and estimated canal seepage were considered for variable inputs.New hydrological insights: Our results showed that ANNs were superior over PLSR in predicting soil salinity, explaining 77% vs. 45% of the variance in soil salinity with root mean square error (RMSE) of 0.12 vs. 0.94 dS/m. In both models, groundwater depth is the most influential variable for soil salinity prediction, with relative contributions of 63% and 65% for PLSR and ANNs, respectively. Though, irrigation water is non-saline river water, it contributes to the rising groundwater table which contains high salinity. Our study demonstrates that proper irrigation management, use of drainage system, and reducing high seepage from the irrigation canal system will sustain the depth of the water table, and simultaneously reduces top-soil salinity accumu-lation and productivity loss in the Rift Valley Region of Awash Basin.

Automated summary

The study focuses on predicting topsoil salinity in irrigated land in the Awash Basin, Ethiopia, using artificial neural networks (ANNs) and partial least squares regression (PLSR). The results show that ANNs outperform PLSR in predicting soil salinity, with better explanatory power and smaller root mean square error.