A generic model of aquatic remote sensing and the incorporation of ecological scenarios using Bayesian statistics

随着人们对全球气候变化影响的日益关切，我们必须监测大面积生态系统的健康状况。卫星或机载传感器的遥感通常被视为实现这一任务最具成本效益的手段。许多遥感研究试图提高地面上可以分辨的东西的分辨率。这是一项复杂的任务，特别是在水生系统中，上覆水柱强烈衰减阳光，因此降低了海床特征颜色的“可分离性”。研究项目通常侧重于特定的生态系统和问题，例如，我们能否区分百慕大珊瑚礁上的活珊瑚和海藻？虽然我们可能会得出答案是“不”，但我们不能将这个答案外推到其他地方。这是因为我们通常不了解结果的确切原因。在这种情况下，是因为南极洲珊瑚和海藻的颜色太相似，还是因为研究地点的水太深或太暗，海浪太强，传感器无法正确观察海床？为了真正理解我们的结果，并将其与其他结果进行比较，我们需要一个遥感在水生环境中如何工作的通用模型。遥感的某些方面已经相当好地理解了，例如光通过水柱的通道。然而，光与结构和光谱复杂的海底（大多数海底都是如此）的相互作用只是最近才被模拟出来。我们使用辐射度方法（在“寻找尼莫”中用于生成珊瑚礁）实现了这种建模。该项目的主要目的是创建和测试一个通用的水生遥感模型（GMARS）。为了创建GMARS，我们将扩展现有的辐射度模型，然后创建空间上逼真的虚拟生态系统，代表两种截然不同的系统类型：波罗的海的藻床和结构复杂的珊瑚礁。这将是第一次在整个水生生态系统中的所有辐射过程都被模拟，并使我们能够测试各种关于遥感仪器局限性的重要假设。我们还将与统计学家合作，以确保参数中的错误和变化可以通过模型传播。重要的是，使用一个正式的统计框架，使我们能够在遥感进一步创新：我们的第二个目标是通过增加其他来源的知识系统被映射，以提高遥感的准确性。遥感算法试图根据某一像素的颜色或纹理来识别该像素。例如，卫星图像的使用者可能试图根据其颜色来区分一块珊瑚礁是珊瑚、沙子还是海藻。然而，我们通常有关于生态系统的先验知识，如果可能的话，应该将其纳入其中。例如，我们可能知道珊瑚礁最近遭受了飓风的袭击，我们有一个生态模型预测珊瑚有70%的机会死亡，并被海藻取代。我们也可能知道，一个像素的沙子在2年内变成珊瑚的可能性很小。使用贝叶斯统计学的一个分支，我们可以正式地将我们的生态预期与遥感仪器的预测相协调。在以前的NERC赠款，我们模拟珊瑚礁种群动态，我们现在将提供一个正式的统计框架，结合联合收割机这些模型的预测与遥感数据。其结果将是改进遥感和生态系统监测。该建议为遥感和生态系统评估提供了两个新的创新：（1）提供一个完全通用的水生系统光模型和（2）一个通用的统计环境，联合收割机结合光谱和生态数据。总之，我们可以确定在遥感过程的每个阶段获取准确数据的价值，这将有助于优先收集实地数据。

With rising concerns about the impacts of global climate change, it is important that we monitor the health of ecosystems over large areas. Remote sensing from satellite or airborne sensors is usually seen as the most cost-effective means of achieving this task. Much remote sensing research attempts to improve the resolution of what can be resolved on the ground. This is a complex task, particularly in aquatic systems where the overlying water column strongly attenuates sunlight and therefore reduces the 'separability' in colour of sea bed features. Research projects usually focus on a specific ecosystem and issue / for example, can we distinguish living corals from seaweeds on a reef in Bermuda? Although we may conclude the answer is 'no', we cannot extrapolate that answer elsewhere. This is because we usually do not understand the precise cause of the result. In this case, are the colours of Bermudian corals and seaweeds too similar or was the water too deep or murky at the study site and the waves too strong to allow the sensor to view the seabed properly? To really understand our results / and compare them to others / we need a generic model of how remote sensing works in an aquatic environment. Some aspects of remote sensing are fairly well understood, such as the passage of light through a water column. However, the interaction of light with a structurally and spectrally complex seabed - as most seabeds are - has only recently been modelled. We achieved this modelling using radiosity methods (which were used to generate reefs in 'Finding Nemo'). The main aim of this project is to create and test a Generic Model of Aquatic Remote Sensing (GMARS). To create GMARS we will extend our existing radiosity model and then create spatially realistic virtual ecosystems that represent two contrasting types of system: algal beds of the Baltic Sea and structurally complex coral reefs. This will be the first time that all radiative processes in an entire aquatic ecosystem have been modelled and allows us to test a variety of important hypotheses about the limitations of remote sensing instruments. We will also collaborate with a statistician to ensure that the errors and variation in parameters can be propagated through the model. Importantly, use of a formal statistical framework allows us to make a further innovation in remote sensing: Our second aim is to improve the accuracy of remote sensing by adding other sources of knowledge about the system being mapped. Remote sensing algorithms attempt to identify a given pixel on the basis of its colour or texture. For example, a user of a satellite image may attempt to discriminate whether a patch of reef is coral, sand or seaweed on the basis of its colour. However, we often have prior knowledge about the ecosystem which should be incorporated if possible. We may know, for example, that the reef was recently struck by a hurricane and we have an ecological model predicts a 70% chance that corals will have died and been replaced by seaweeds. We may also know that it is highly unlikely that a pixel of sand will turn into a coral within a period of say 2 years. Using a branch of Bayesian statistics, we can formally reconcile our ecological expectations with the predictions made from a remote sensing instrument. In a previous NERC grant, we modelled coral reef population dynamics and we will now provide a formal statistical framework to combine these model predictions with those of remotely-sensed data. The result will be improved remote sensing and ecosystem monitoring. This proposal provides two new innovations for remote sensing and ecosystem assessment: (1) provision of a fully generic model of light in aquatic systems and (2) a generic statistical environment to combine both spectral and ecological data. Taken together, we can identify the value of acquiring accurate data at each stage of the remote sensing process which will help prioritise the collection of field data.