In recent years, the application of artificial intelligence and deep learning in agriculture has gained significant attention due to its ability to improve crop monitoring and field analysis. However, one of the major challenges in developing effective deep learning models is the limited availability of large and diverse datasets. Agricultural datasets are often small and do not represent real-world variations such as changes in lighting, soil conditions, and crop growth stages. Traditional data augmentation techniques such as rotation, flipping, and scaling help to increase dataset size but fail to capture realistic variations. To overcome this limitation, this work proposes a practical approach for generating synthetic agricultural images using segmentation-based augmentation. Crop and weed regions are extracted from field images and combined with different soil backgrounds to create new training samples. These synthetic images improve dataset diversity and help deep learning models learn more effectively. The augmented dataset is used to train models such as CNN, ResNet50, and YOLOv8 for classification and detection tasks. Experimental results show improved accuracy, reduced overfitting, and better generalization. The proposed approach is simple, efficient, and suitable for real-world agricultural applications where data availability is limited.