This work proposes a 3D U-Net architecture for the automatic segmentation of soybean fields using multitemporal and multispectral satellite imagery. A dataset was generated through the Google Earth Engine (GEE), combining spectral bands, vegetation indices (Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI)), and the MapBiomas soybean mask for the 2021/2022 growing season in southern Brazil. The use of vegetation indices is supported by their proven effectiveness in agricultural monitoring [1]. To address the strong class imbalance between soybean and non-soybean regions, a stratified balancing strategy was applied, ensuring a more uniform distribution across different levels of crop coverage. The proposed model integrates temporal information across five months of the crop cycle, enabling the extraction of both spatial and temporal patterns, similar to other works that explore multitemporal deep learning architectures [2]. A hybrid loss function combining balanced Binary Cross-Entropy (BCE) and the Dice coefficient was employed to improve segmentation accuracy, especially in regions with low soybean presence. Experimental results demonstrate that the model achieves stable learning, with early stopping preventing overfitting and the best validation performance occurring at epoch 13. Visual analyses confirm strong spatial consistency between predictions and reference masks. The study highlights the potential of 3D convolutional architectures for large-scale automated crop mapping [3] and suggests future improvements by incorporating additional seasons and broader geographic coverage.