Models that can generate realistic Earth surface patterns are important both for geomorphological applications and as prior models for underdetermined inverse problems. Generative machine learning methods such as GANs and the increasing availability of large remote sensing data sets represents an exciting combination for this purpose. Several studies show promising results for GANs trained on artificial data sets in geostatistics, but it is necessary to further quantify how well such models reproduce and generalize real data. The conditioning ability of GANs is often evaluated based on output which originates from a trained generator. In reality, geophysical data necessarily arises from elsewhere. Here, we use more realistic training data than in previous studies and evaluate performance using an extensive set of metrics and real images outside the training data set. The data set consists of multispectral satellite imagery of 38 large river deltas, a type of Earth surface pattern which is limited in number. The channel network is used to create training images with four sedimentary facies, which are subsequently used to train a Wasserstein GAN of deltaic 2D patterns. GANs successfully reproduce all training data characteristics and produce manifold the number of combinations with respect to the training data. However, there does not seem to be an infinite number of discrete combinations of facies, and the posterior landscapes are not well‐shaped for efficient exploration in the presence of so‐called hard data. Thus, GANs should have many exciting applications in geosciences, but it will depend on the type of measurement data.

中文翻译：

---

利用卫星图像数据集上训练的GAN模拟河流格局

可以生成逼真的地球表面模式的模型对于地貌应用以及对于欠定反演问题的现有模型都非常重要。诸如GAN之类的生成式机器学习方法以及大型遥感数据集的日益普及，代表了一种令人兴奋的组合。多项研究表明，对在地统计学中的人工数据集进行训练的GAN具有令人鼓舞的结果，但有必要进一步量化此类模型对真实数据的再现和推广程度。经常根据来自受过训练的发电机的输出来评估GAN的调节能力。实际上，地球物理数据必然来自其他地方。这里，与以前的研究相比，我们使用了更现实的训练数据，并使用训练数据集之外的大量指标和真实图像评估性能。数据集由38个大河三角洲的多光谱卫星图像组成，这是数量有限的一种地球表面模式。通道网络用于创建具有四个沉积相的训练图像，随后用于训练具有二维2D模式的Wasserstein GAN。GAN成功地重现了所有训练数据的特征，并针对训练数据产生了多种组合。但是，似乎没有无数个离散的相组合，并且在存在所谓的硬数据的情况下，后部景观的形状也不适合有效勘探。因此，