A well designed observing network is vital to improve our understanding of the oceans and to obtain better predictions of the future. As autonomous marine technology develops, the potential for deploying large autonomous observing systems becomes feasible. Though there are many design considerations to take into account (according to the target data use cases), a fundamental requirement is to take observations that capture the variability at the appropriate length scales. In doing so, a balance must be struck between the limited observation resources available and how well they are able to represent different areas of the ocean. In this paper we present and evaluate a new method to aid decision makers in designing near-optimal observing networks. The method uses ideas from set theory to recommend an irregular network of observations which provides a guaranteed level of representation (correlation) across a domain. We show that our method places more observations in areas with smaller characteristic length scales and vice versa, as desired. We compare the method to two other grid types: regular and randomly allocated observation locations. Our new method is able to provide comparable average representation of data across the domain, whilst efficiently targeting resource to regions with shorter length scale and thereby elevating the minimum skill baseline, compared to the other two grid types. The method is also able to provide a network that represents up to 15% more of the domain area. Assessing error metrics such as Root Mean Square Error and correlation shows that our method is able to reconstruct data more consistently across all length scales, especially at smaller scales where we see RMSE 2-3 times lower and correlations of over 0.2 higher. We provide an additional discussion on the variability inherent in such methods as well as practical advice for the user. We show that considerations must be made based on time filtering, seasonality, depth and horizontal resolution.

一个设计良好的观测网络对于提高我们对海洋的了解和更好地预测未来至关重要。随着自主海洋技术的发展，部署大型自主观测系统的潜力变得可行。尽管有许多设计考虑因素需要考虑（根据目标数据用例），但基本要求是进行观察，以捕捉适当长度尺度上的可变性。在此过程中，必须在可用的有限观测资源与它们代表海洋不同区域的能力之间取得平衡。在本文中，我们介绍并评估了一种新方法，以帮助决策者设计接近最佳的观测网络。该方法使用集合论的思想来推荐一个不规则的观察网络，该网络提供跨域的保证水平的表示（相关性）。我们表明，我们的方法根据需要在具有较小特征长度尺度的区域放置更多观测值，反之亦然。我们将该方法与其他两种网格类型进行比较：常规和随机分配的观察位置。与其他两种网格类型相比，我们的新方法能够提供跨域数据的可比较平均表示，同时有效地将资源定位到长度尺度较短的区域，从而提高最低技能基线。该方法还能够提供一个代表高达 15% 以上的域区域的网络。评估均方根误差和相关性等误差指标表明，我们的方法能够在所有长度尺度上更一致地重建数据，尤其是在较小的尺度上，我们看到 RMSE 低 2-3 倍，相关性高出 0.2 以上。我们对此类方法固有的可变性进行了额外讨论，并为用户提供了实用建议。我们表明，必须根据时间过滤、季节性、深度和水平分辨率进行考虑。