Optimizing cotton green water footprint prediction using hybrid machine learning algorithms: a case study of Al-Gezira state, Sudan

Abstract Water scarcity and climate change pose significant challenges for Sudan, leading to considerable migration. A total of 1 million hectares of arable land are irrigated, while 6.7 million hectares employ semi-mechanized rainfed agricultural practices. In contrast, a significant 9 million hectares depend solely on conventional rainfed techniques. GWFP deals with precipitation stored in the soil as moisture and consumed in biomass production, as agricultural products are usually irrigated with rainwater and thus more dependent on green water sources. Calculating the green water footprint is important for developing sustainable agricultural practices and effectively managing water resources. The accurate estimation of the GWFP value is very important in economics as an approach to foster the virtual green water trade and improve human well-being. This research aims to assess the efficacy of machine learning models in predicting the green water footprint (GWFP) of cotton within the framework of climate change. By examining a range of input variables, including climatic conditions, agricultural data, and remote sensing indices, the study explores their impacts on cotton cultivation over the time frame from 2001 to 2020. A total of seven models were implemented, comprising random forest (RF), Extreme Gradient Boosting (XGBoost), and support vector regressor (SVR), along with hybrid combinations such as RF-XGB, RF-SVR, XGB-SVR, and RF-XGB-SVR, across five scenarios (Sc) incorporating diverse variable combinations utilized throughout the investigation. The maximum and minimum RMSE values varied between 31.35 m 3 t −1 and 166.37 m 3 t −1 , based on the RF-XGB-SVR hybrid model and the RF model, respectively, under Sc5 (Peeff, and Tmax). The highest R 2 values were achieved with hybrid ML models, whether double or triple, across all scenarios, reaching values of 1.0 or 0.99. The lowest R 2 value, recorded at 0.0676, was noted under SVR and Sc3, followed closely by XGB and Sc3 with a value of 0.0767. The box plot for GWFP of cotton indicated that the XGB-SVR and Sc3 exhibited the lowest interquartile range (IQR) at 0.047, succeeded by the RF-XGB-SVR model with Sc3 at a value of 0.052; however, the XGB-SVR hybrid model displayed the highest IQR in Sc5 at 0.098. The research concludes that hybrid models outperformed single models in forecasting cotton GWFP. Furthermore, remote sensing indices showed a negligible positive impact on GWFP prediction, with Sc3 yielding the lowest statistical results across all models. The study recommends the employment of hybrid models to reduce the error term in predicting cotton GWFP.