SatSynth is a method that uses diffusion models to generate synthetic image-mask pairs for satellite semantic segmentation. The approach leverages generative image diffusion to create new training data instances, which are then used to improve semantic segmentation performance. The method addresses the challenge of limited annotated data in earth observation tasks by learning the joint data distribution of images and labels. The generated pairs are of high quality and diverse, which is crucial for earth observation where semantic classes vary in scale and occurrence. The synthetic data is used as a form of data augmentation to train downstream segmentation models. Experiments show that integrating generated samples significantly improves performance on satellite segmentation benchmarks compared to baselines and when training only on original data. The method is evaluated on three satellite datasets (iSAID, LoveDA, OpenEarthMap) and demonstrates improvements in segmentation accuracy. The approach also includes techniques for image super-resolution to enhance the resolution of generated images. The results show that the proposed method outperforms existing generative models in terms of visual quality and segmentation accuracy. The work highlights the potential of diffusion models for data synthesis in domains with scarce and costly ground truth labels.SatSynth is a method that uses diffusion models to generate synthetic image-mask pairs for satellite semantic segmentation. The approach leverages generative image diffusion to create new training data instances, which are then used to improve semantic segmentation performance. The method addresses the challenge of limited annotated data in earth observation tasks by learning the joint data distribution of images and labels. The generated pairs are of high quality and diverse, which is crucial for earth observation where semantic classes vary in scale and occurrence. The synthetic data is used as a form of data augmentation to train downstream segmentation models. Experiments show that integrating generated samples significantly improves performance on satellite segmentation benchmarks compared to baselines and when training only on original data. The method is evaluated on three satellite datasets (iSAID, LoveDA, OpenEarthMap) and demonstrates improvements in segmentation accuracy. The approach also includes techniques for image super-resolution to enhance the resolution of generated images. The results show that the proposed method outperforms existing generative models in terms of visual quality and segmentation accuracy. The work highlights the potential of diffusion models for data synthesis in domains with scarce and costly ground truth labels.