Generative adversarial network augmentation for solving the training data imbalance problem in crop classification

Deep learning models offer great potential for advancing land monitoring using satellite data. However, they face challenges due to imbalanced real-world data distributions of land cover and crop types, hindering scalability and transferability. This letter presents a novel data augmentation method employing Generative Adversarial Neural Networks (GANs) with pixel-to-pixel transformation (pix2pix). This approach generates realistic synthetic satellite images with artificial ground truth masks, even for rare crop class distributions. It enables the creation of additional minority class samples, enhancing control over training data balance and outperforming traditional augmentation methods. Implementing this method improved the overall map accuracy (OA) and intersection over union (IoU) by 1.5% and 2.1%, while average crop type classes' user accuracy (UA) and producer accuracies (PA), as well as IoU, were improved by 11.2%, 6.4% and 10.2%.