In this study, a convolutional neural network (CNN)-based deep learning model was developed to predict durum wheat yield by integrating RGB and multispectral UAV imagery, ground-based sensor measurements (SPAD, NDVI, LAI, and plant height), and climatic parameters under semi-arid conditions. The dataset was designed as a multi-source, multi-stage, and multi-year structure, comprising measurements collected across six phenological growth stages during the 2023 and 2024 growing seasons. Principal component analysis (PCA) indicated that approximately 69% of the total variance was explained by the first two components, with SPAD, NDVI, LAI, and plant height identified as the most influential variables in explaining yield variability. The CNN model achieved high predictive accuracy in both stage-based and year-based evaluations, with R2 values ranging from 0.982 to 0.994, RMSE between 0.15 and 0.24 kg ha- 1, and MAE between 0.11 and 0.19 kg ha- 1. The highest performance was obtained during the heading and grain-filling stages. Overall, the results demonstrate that integrating UAV imagery, physiological sensor indicators, and climatic variables within a multi-source, multi-stage, multi-year deep learning framework substantially improves yield prediction accuracy compared with single-source approaches. This study presents a high-performance CNN architecture for yield forecasting and provides a robust foundation for generalizable and effective decision-support systems in precision agriculture.