Application of deep learning for sustainable water management by estimating reference crop evapotranspiration in sub-humid climatic conditions

The sustainability of water resources depends on a very accurate Reference Evapotranspiration (ETo) estimation of good irrigation planning and scheduling. Achieving sustainability uses modern technologies like Deep Learning (DL). This study will contribute to filling the gap in the artificial intelligence uses in agriculture that related to using DL methods in estimating daily ETo. The study aims to determine the best DL architecture type in estimating daily ETo by testing MLP, 1D-CNN, LSTM, Bi-LSTM, and Hybrid (CNN-LSTM) architectures, defining the best selection of meteorological parameters to estimate a high accuracy of ETo with limited parameters. Finally, it aims to determine the minimum amount of data used in training DL models to achieve high accuracy in ETo estimation and which the best DL architecture uses a small amount of data. In this study, daily ETo was estimated by 22 data combinations created from meteorology parameters, and each of the data combinations was tested by K-Fold Cross Validation on 93 different DL architecture types. Daily ETo values estimated by the FAO Penman-Monteith method were used when comparing results from DL models. The evaluation of DL architectures' performances in estimating daily ETo showed that the highest and balanced performance DL architectures is hybrid, and 1D-CNN architecture comes after it. The lowest and unstable performance one is MLP architectures. As for LSTM and Bi-LSTM architecture performances are well; there is no significant difference in performance between each other, but Bi-LSTM is slightly better than LSTM. The selection of limited parameters to predict high-efficiency ETo with any DL architecture; It has been found that average or minimum-maximum temperature and sunshine duration should be preferred over other parameters. Accordingly, the best data combinations were created from one, two, and three parameters for estimating ETo: temperature, temperature- sunshine duration, temperature- sunshine duration-wind speed. Finally, in the five types of DL architectures, changing the amount of data used in training the models between 1 and 26 years did not significantly affect the estimation of 5 years of daily ETo in addition, Hybrid architecture outperformed other DL architectures when using a fewer amount of data in estimating ETo.