Reference evapotranspiration (ETo) constitutes a fundamental component in water resource management and irrigation scheduling, playing an important role in agricultural planning, hydrological modeling, and climate change impact assessment. Accurate estimation of ETo primarily depends upon the availability of comprehensive meteorological parameters, including maximum and minimum temperature, relative humidity, wind speed, and solar radiation, which are required for the implementation of the FAO-56 Penman-Monteith method, the internationally recognized standard for ETo calculation. However, ground-based meteorological data often encounter significant limitations in many regions due to sparse and heterogeneously distributed station networks, high maintenance costs, and limited accessibility, particularly in developing countries and areas with complex topography. These constraints lead to considerable uncertainty in ETo estimation and consequently create difficulties in sustainable water resource management, irrigation scheduling, and crop water requirement assessment. In response to these challenges, remote sensing technology has emerged as a promising alternative, offering spatially continuous, temporally consistent, and cost-effective data coverage at regional and global scales.

This study was conducted with the primary objective of assessing the feasibility of replacing ground-based meteorological data with remote sensing parameters for daily ETo estimation using the FAO-56 Penman-Monteith method. The research employed data from 15 weather stations in East Azerbaijan Province in northwestern Iran, covering the period from 2000 to 2023. This region was selected due to its semi-arid to cold semi-arid climate, significant topographical heterogeneity ranging from approximately 700 meters to over 1800 meters above sea level, and its agricultural importance, making it an ideal testbed for evaluating the potential of remote sensing data in complex environmental conditions. The ground-based meteorological data, obtained from the Islamic Republic of Iran Meteorological Organization (IRIMO), included daily maximum temperature (Tmax), minimum temperature (Tmin), mean relative humidity (RH), wind speed at 2-meter height (Ws), incoming solar radiation (Rs), and extraterrestrial radiation (Ra). These data were complemented by MODIS satellite products, including land surface temperature (LST) from both daytime and nighttime observations, solar radiation (DSR), albedo, Normalized Difference Vegetation Index (NDVI), and solar zenith angle (SZA), as well as soil moisture data obtained from the GLDAS satellite product, all extracted through the Google Earth Engine platform.

Four machine learning models were developed and evaluated with six distinct input scenarios, ranging from minimal ground-based variables to a full combination of satellite and ground-based data. The models included Gaussian Process Regression (GPR), a Bayesian non-parametric method capable of estimating prediction uncertainty; LightGBM, an efficient gradient boosting decision tree algorithm suitable for large-scale regression problems; Temporal Convolutional Network (TCN), a deep learning architecture based on causal and dilated convolutions with residual connections; and a hybrid TCN-GRU architecture designed to combine the multi-scale feature extraction capabilities of TCN with the long-term dependency modeling strengths of Gated Recurrent Units (GRU). The input scenarios were carefully designed to evaluate the incremental contribution of different data sources: IC1 used only ground-based variables (Tmax, Tmin, RH); IC2 and IC3 employed limited satellite variables; IC4 incorporated complete satellite-derived variables; IC5 combined complete satellite data with Tmax and RH; and IC6 combined satellite data with Tmin and Ws.

The results demonstrated that the TCN-GRU model provided the most accurate performance under the hybrid scenario IC5, which incorporated all satellite-derived variables alongside maximum temperature and relative humidity. This configuration yielded a coefficient of determination (NSE) of 0.975, a mean absolute error (MAE) of 0.29 mm/day, and a normalized root-mean-square error (NRMSE) of 7.6%, indicating excellent agreement between estimated and observed ETo values. In contrast, the exclusively satellite-based scenarios (IC2–IC4) exhibited a substantial decline in estimation accuracy, with the TCN-GRU model achieving NSE values below 0.79 in IC2 and NRMSE values exceeding 10.6% even in the most comprehensive satellite-only scenario (IC4), highlighting the inherent limitations of remote sensing data when used in isolation for ETo estimation.

The Diebold–Mariano test confirmed that the superiority of the TCN-GRU model over GPR and LightGBM was statistically significant across all scenarios at the 0.01 significance level, while its advantage over the standalone TCN was only significant under the IC5 scenario at the 0.05 level. This finding underscores the value of the hybrid architecture in effectively modeling temporal dependencies and spatial patterns in ETo estimation. Pearson correlation analysis further revealed that maximum temperature (r = 0.93) and solar radiation (r = 0.91) showed the strongest positive correlations with ETo, which is consistent with the physical principles of evapotranspiration, where temperature and energy availability are primary driving factors. Conversely, solar zenith angle (r = -0.88) and relative humidity (r = -0.73) showed the strongest negative correlations with ETo, indicating their role as limiting factors through their influence on atmospheric demand and energy receipt at the surface.

Overall, although remote sensing data cannot fully substitute for ground-based meteorological measurements, their integration with a minimal set of ground variables, specifically maximum temperature and relative humidity, can yield acceptable accuracy (NRMSE < 8%) for operational applications. The proposed TCN-GRU model, with its hybrid convolutional-recurrent architecture, effectively captures both short-term and long-term dependencies in meteorological time series and is recommended as an efficient tool for ETo estimation in regions with limited ground-based data availability.