Molajou, AmirAram, Saeed Hedayati2026-04-212026-04-2120262772-375510.1016/j.atech.2026.101934https://hdl.handle.net/123456789/9117https://doi.org/10.1016/j.atech.2026.101934Context: Accurate wheat yield prediction (WYP) is a cornerstone of global food security. This is especially critical for nations like Iran, which confronts significant environmental challenges, including water scarcity, extreme climate events, and soil degradation. The country's diverse range of climates further complicates agricultural management, making reliable yield forecasting a complex but essential task. Objectives: A significant gap in current literature is the lack of model validation across different climatic zones, which restricts the transferability and widespread adoption of existing prediction systems. Also, the operational collection of field-scale data is often limited due to the high costs, and relatively few studies have been conducted at this scale. Additionally, the WYP models are frequently compromised by persistent data quality issues, such as cloud cover in satellite imagery. Therefore, this study aimed to develop and validate a robust WYP framework specifically designed for the diverse agro-climatic zones of Iran. The objective was to create a scalable and reliable system that directly addresses the limitations of cross-climate validation and data imperfections. Methods: A comprehensive six-year dataset spanning from 2019 to 2024 was first created. This was achieved by fusing Sentinel-2 satellite imagery with ERA5 temperature data. The model inputs consisted of Sentinel-2 Level2A surface reflectance from NIR (Band 8) and SWIR bands (Bands 11 and 12), along with ERA5 temperature data including air temperature at 2 meters and soil temperatures at depths ranging from 0 to 289 cm. To address the issue of cloud cover, two distinct data processing strategies were developed and evaluated. The first was a timeseries (TS) approach that utilized monthly maximum band values. The second involved a cumulative summation of these monthly values. A comparative analysis was then conducted between two machine learning models: the widely used XGBoost algorithm and a Bayesian Neural Network (BNN). The BNN was specifically selected for its ability to quantify the uncertainty associated with its predictions. The performance of these models was systematically assessed across various agro-climatic zones in Iran. Results and conclusions: It was found that the BNN consistently outperformed the XGBoost model in all tested climate zones. Notably, the highest accuracy was achieved in the most challenging arid environments. The superiority of the TS data processing method was also clearly established. For the TS scenario, a normalized root mean square error (nRMSE) range of 0.21-0.46 was recorded. In contrast, the summation scenario yielded a higher nRMSE range of 0.26-0.54. Significance: Present research demonstrates a scalable, uncertainty-aware methodology that advances precision agriculture and strengthens climate adaptation strategies, providing a validated pathway toward more reliable WYP systems.eninfo:eu-repo/semantics/openAccessRemote SensingDeep LearningUncertainty QuantificationWheat Yield PredictionField-scaleArticle