Predicting blue water footprint and green water footprint of orange in Nile Delta based on single, hybrid, and stacking hybridization machine learning algorithms under diverse agro-climatic conditions

This study aims to investigate the effects of climate change on the Blue water footprint (BWFP) and Green water footprint (GWFP) of oranges in the El-Beheira and Al-Sharkia governorates in the Nile Delta using data from 2013 to 2022. Various machine learning (ML) models, including LASSO, RF, XGB, and CATBOOST, as well as hybrid models like XGB-RF, RF-LASSO-CAT-LASSO, XGB-LASSO, XGB-CAT, XGB-CAT-LASSO, XGB-RF-LASSO, and stacked ensemble models, were employed to assess the impact of climate change in Egypt and determine the most accurate model for predicting BWFP and GWFP of orange crops. The models were tested under different scenarios (Sc) involving climate parameters, crop parameters, and remote sensing indices (Normalized Difference Vegetative Index (NDVI), Normalized Difference Moisture Index (NDMI), Enhanced vegetation index (EVI), Soil-adjusted Vegetation Index (SAVI), Green Chlorophyll Index (GCI), and Land Surface Temperature (LST)). The CatBoost-RF hybrid model in Sc5 (Peeff, Tmax) showed the highest accuracy of 1.00 for BWFP, while the LASSO model in Sc3 had the lowest accuracy of 0.846. For GWFP, all hybrid models achieved a maximum accuracy of 1.0. The R2 value for predicting BWFP was highest at 0.98 under Sc2 using RF-LASSO hybrid models and lowest at 0.63 under Sc3 with RF. As for GWFP, the highest R2 value of 1.0 was obtained under Sc1 with XGB-CatBoost-LASSO, while the lowest value of 0.07 was seen under Sc3 with XGB-RF-LASSO. The lowest RMSE value of 0.15 was observed with the RF-LASSO hybrid model, while the highest value of 0.31 was seen with the LASSO model in Sc2 for BWFP. For GWFP, the minimum and maximum RMSE values were 0.01 and 0.06 under XGB-RF-LASSO and CatBoost models, respectively, in Sc4 (Peeff, Tmax, Tmin). The study suggests utilizing hybrid models, either double or triple, especially for RF-LASSO and XGB-CAT-LASSO, to effectively predict the BWFP and GWFP of oranges in arid regions.