• A deep learning-based CNN fusion method was developed for ETo estimation. • Multi source remote sensing data were fused through different pixel-level fusion methods. • Gaussian noise escalated error magnitudes of the fusion-based and direct methods. • RF-CNN provided the most accurate results under clean and noisy data conditions. • SHAP analysis identified the features importance level. When using satellite-based data for ET o estimations, fusion techniques might improve the accuracy of modeling performance by integrating multiple source predictors and reducing uncertainty. However, most of the existing studies have compared the fusion-based estimations with standalone-source-based outputs (where data from a single source is used for ET o estimation). Nevertheless, there is still an opportunity to evaluate the fusion-based estimation of ET o values when different multi-source data are used in the simulation procedure, which has not been addressed yet. Furthermore, the performance of fusion-based models under noisy input conditions, as well as the impact of each input variable on simulating ET o using these methods, remains unassessed. The present study aimed at developing a deep learning-based fusion method for integrating the satellite-based high-resolution data from different sources, namely MODIS, ERA5, and CHIRPS, through a convolutional neural network (CNN) fusion method. Data from two humid and arid locations were utilized for assessing the suggested methodology. A comparison was made between CNN-based fusion with other pixel-level fusion methods (Weighted Average, Principal Component Analysis, Kalman Filter, and Autoencoder). The capabilities of the fusion methods were examined through coupling with the random forest (RF) model for estimating daily ET o values. All the applied models were evaluated under clean- and noisy-data conditions through importing Gaussian noise to the predictors. The capability and stability of the models were evaluated using statistical indices as well as the residual and sensitivity analyses. SHAP theory was used for detecting the most influential parameters on the simulation procedure. The results revealed that the RF model coupled with the CNN-fusion method outperformed the rest of the models from both performance accuracy and model stability viewpoints under clean and noisy data conditions. Specifically, for clean data, the RF-CNN model provided R 2 values of 0.990 and 0.997, and RMSE values of 0.177 mm d -1 and 0.092 mm d -1 for the humid and arid regions, respectively, which were further validated using Monte-Carlo analysis.
Sadeghzadeh et al. (Sun,) studied this question.