<p>Air temperature is a critical climatic variable for agricultural production, influenced by multiple environmental factors and exhibiting pronounced spatial variability. In large agricultural regions of Brazil, its characterization is constrained by the low density and uneven distribution of meteorological stations. Despite the increasing applications of deep learning artificial neural networks, the importance of variable selection and model complexity at the monthly scale remains unclear in the Central-West region of Brazil. This study evaluated the applicability of deep learning models to estimate monthly maximum (Tmax), mean (Tm), and minimum (Tmin) air temperatures using long-term surface meteorological data from 81 weather stations covering the period 1990–2020. Nine explanatory variable scenarios (S1–S9), combining geographic and meteorological predictors, were assessed using deep learning architectures with different hidden layer configurations (M1, M2, and M3), and model performance was evaluated using training, validation, and test datasets. The results indicate that scenarios S1, S2, S3, S4, and S6 yielded more consistent estimates across months and network configurations, with a larger proportion of models achieving R² ≥ 70%. Model M3 for Tmax and Tm and model M2 for Tmin were associated with satisfactory performance. The results suggest that simpler and physically consistent input combinations (e.g., S6, based on altitude and latitude) can estimate spatiotemporal temperatures while maintaining model stability. The study highlights the importance of selecting explanatory variables over increasing model complexity, indicating that physical consistency can support accurate monthly estimates of air temperature, providing an efficient alternative for regions with limited meteorological data.</p>