Application of artificial neural networks in the design of drainage systems in data-poor areas.

Drainage has been identified as an often-neglected component of irrigated agriculture in arid and semi-arid areas. Even though it is accepted that drainage is often necessary to prevent waterlogging and salinity impacting productivity in irrigated agriculture, it is typically ignored when planning future irrigation schemes. Only 5 – 10% of the total irrigated land in Least Developed Countries (LDCs) that requires drainage is currently drained (compared to 25 – 30% in developed countries). This is partly due to a fundamental lack of spatially and temporally coherent datasets containing key input parameters for drainage models, local expertise and the high cost of drainage installation. Drainage simulation models can provide reliable predictions of multi-component systems to evaluate drainage system design over long periods (1 – 100 years). This study evaluated existing drainage simulation models (i.e. DRAINMOD, SWAP, ADAPT, RZWQM2, EPIC, WaSim and HYDRUS-1D) for their suitability to be applied in data-poor arid and semi-arid regions. Based on a selection criteria, the most applicable model for drainage design in arid and semi-arid areas was DRAINMOD. DRAINMOD, an agricultural drainage simulation model, is a versatile and readily available model that can be used to evaluate alternative drainage system designs. DRAINMOD requires several key inputs, including saturated hydraulic conductivity (Ksat), reference evapotranspiration (ET0) and the Electrical Conductivity of a saturated soil Extract (ECe). In LDCs, measuring these parameters is expensive and time-consuming. In addition, existing historic datasets are often spatially and temporally limited. Therefore, indirect approaches are needed to overcome incomplete data records that restrict drainage designs. This thesis evaluates the feasibility of applying indirect methods, with a focus on developing and validating the use of artificial neural networks (ANNs) using available historic measured datasets. The study data draws on the drainage design for Hammam Agricultural Project (HAP) and Eshkeda Agricultural Project (EAP), located in the south of Libya, north of the Sahara Desert. Soil texture, bulk density, field capacity, and wilting point were used to develop ANNs to predict Ksat which were significantly more accurate compared to widely adopted Pedotransfer functions (PTFs) such as Rosetta3. To calculate the daily ET0, average monthly maximum and minimum air temperature were used to develop ANNs. Arithmetic Averaging of Neighbouring Stations (AANS), MODAWEC and Era5-Land were among the indirect methods applied to predict ET0. Landsat 5 Surface Reflectance bands and the derived salinity indices were applied to develop ANNs to estimate ECe. The accuracy of the predicted values of Ksat, ET0 and ECe were evaluated by using statistical parameters such as coefficient of determination (R²), mean square error (MSE), and root mean square error (RMSE). The predicted Ksat and ET0 values were input to DRAINMOD to design drainage systems in EAP and HAP as compared to the optimum design based on measured data. The design focused on how accurately the predicted values were able to estimate drain spacing, relative yield, irrigation depth, and drainage discharge. The key findings showed that the accuracy of predicting Ksat greatly impacted predicting the optimum drain spacing and the associated relative yield. Accurate prediction of the optimum spacing between drains will reduce the overall cost by ensuring that the drains are not spaced too closely, but also lowers the risk of raising the water table and negatively impacting the yield by preventing the drains being installed on too wider a spacing. In addition, precisely predicting ET0 is essential to quantify the irrigation water requirement and drainage discharge. Finally, predicting soil salinity using remote sensing data can be used as an early warning tool to monitor irrigated lands affected by salinity, evaluate the performance of existing drainage systems, and indicate areas that need improvement. Future research recommendations identified by this research include the need for (1) critical evaluation of the accuracy of using ANNs and other machine learning approaches to predict other input parameters required for drainage design such as the water retention curve, depth of impermeable layer, hourly or daily rainfall, and initial water table depth. (2) development and validation of ANNs and other machine learning approaches that can predict Ksat, ET0, and ECe on a national level (Libya) and/or regional level (Middle East and North Africa) to overcome the challenge of incomplete data records that restrict drainage designs.