A coherent portrayal of global bathymetry requires that depths are inferred between sparsely distributed direct depth measurements. Depths can be interpolated in the gaps using alternate information such as satellite-derived gravity and a mapping from gravity to depth. We designed and trained a neural network on a collection of 50 million depth soundings to predict bathymetry globally using gravity anomalies. We find the best result is achieved by pre-filtering depth and gravity in accordance with isostatic admittance theory described in previous predicted depth studies. When training the model, if the training and testing split is a random partition at the same resolution as the data, the training and testing sets will not be independent, and model misfit is underestimated. We solve this problem by partitioning the training and testing set with geographic bins. Our final predicted depth model improves on old predicted depth model RMSE by 16%, from 165 to 138 m. Among constrained grid cells, 80% of the predicted values are within 128 m of the true value. Improvements to this model will continue with additional depth measurements, but predictions at higher spatial resolution, being limited by upward continuation of gravity, should not be attempted with this method.

中文翻译：

---

使用神经网络进行全球预测水深测量

全球测深的连贯描述需要在稀疏分布的直接深度测量之间推断深度。可以使用替代信息（例如卫星衍生的重力和从重力到深度的映射）将深度插值到间隙中。我们设计并训练了一个神经网络，收集了 5000 万次深度探测数据，以利用重力异常预测全球测深。我们发现最好的结果是根据先前预测深度研究中描述的等静压导纳理论预过滤深度和重力来实现的。在训练模型时，如果训练和测试分割是与数据相同分辨率的随机分割，则训练集和测试集将不独立，并且模型失配被低估。我们通过使用地理区间对训练和测试集进行分区来解决这个问题。我们最终的预测深度模型比旧的预测深度模型 RMSE 提高了 16%，从 165 m 提高到 138 m。在受约束的网格单元中，80% 的预测值与真实值的偏差在 128 m 以内。对该模型的改进将继续进行额外的深度测量，但由于重力向上延续的限制，不应尝试使用此方法进行更高空间分辨率的预测。