Estimating actual crop evapotranspiration by using satellite images coupled with hybrid deep learning-based models in potato fields

Estimating actual crop evapotranspiration (ET c act ) is a fundamental requirement for effective crop water management, aiming to achieve precision irrigation amidst changing environmental conditions. Despite the introduction of various methods for estimating ET c act values, these methods are still associated with several challenges and limitations. Motivated by the robustness of deep learning models, this study employed two hybrid models that integrate Convolution Neural Network with either Random Forests (CNN-RF) or Support Vector Machine (CNN-SVR) to estimate potato ET c act using a limited set of input features. Three models were configured and compared for each CNN-RF (CNN-RF 1 , CNNRF 2 , CNNRF 3 ) and CNN-SVM (CNN-SVM 1 , CNN-SVM 2 , CNN-SVM 3 ), by using different combinations of variable input features derived from meteorological data (air temperature (Ta), vapour pressure deficit (VPD), net radiation (Rn)) and MODIS satellite data (land surface temperature (LST), fraction of photosynthetically active radiation (Fpar), leaf area index (LAI)). The results demonstrated the outstanding performance of the CNN-RF models, showcasing their superiority over the CNN-SVM models. Notably CNN-RF 1 model yielded the best results, achieving a normalized root mean squared error (nRMSE) of 11.7 and 31.5 %; Nash–Sutcliffe Efficiency (NSE) of 0.97 and 0.80; Correlation coefficient (CC) of 0.99 and 0.91 at training and testing phases respectively. While, both models exhibited lower performance when using remote sensing satellite-based inputs, CNN-RF 2 , managed to produce satisfactory estimates of nRMSE = 16.7, 40.9 %; NSE = 0.95, 0.66; CC = 0.98, 0.813; Bias = −0.41, −4.377 W/m 2 during training and testing respectively. The ET c act of potato showed to have high correlation with LST (0.74–0.84) and using LST as proxy for Ta, resulted in improved model estimates compared to using satellite data exclusively. These findings suggest that, in situations where climatic data is limited, remote sensing can serve as a viable alternative for modelling potato ET c act when coupled with CNN-RF, and further improvement can be achieved by incorporating meteorological data fusion. • Estimation of ET c act is a fundamental requirement of crop water management. • Deep learning based Random Forest models demonstrated to be superior when estimating potato ET c act . • Satellite remote sensing data is a viable option when estimating potato ET c act .