Modelling spatial distribution and change from wildlife survey data

减少生物多样性损失是《生物多样性公约》(CBD)2020年的一个关键目标，量化损失是管理生物多样性的关键。这涉及到估计野生种群的大小和分布，这在统计学上具有挑战性--只使用检测到的动物(通常是种群的一小部分)，必须推断出没有检测到的动物的丰度和分布。自然系统总是有空间结构，监测和了解是什么驱动了栖息地的利用、空间分布和空间分布的变化，这是理解和预测自然或人类诱导的扰动对自然系统的影响的核心。这是困难的，因为动植物的空间结构往往是复杂的，涉及空间趋势、空间随机性和空间相关性。不能适应空间分布的所有这些方面的空间模型的拟合可能会导致关于空间分布的驱动因素和分布变化的非常误导的结论。特别是，对随机性和相关性的建模不当可能会导致不正确的推断和误导性的预测。虽然现实中复杂的空间模型已经存在了一段时间，但直到最近，对这种模型进行拟合的方法还太慢，无法发挥作用。随着集成嵌套拉普拉斯近似(INLA)方法的出现，这种情况不再是这样的，因此，该方法的使用迅速增长，实现该方法的软件需求很大。然而，目前还没有方法或软件(INLA或其他)来将实际复杂的空间模型与从其中检测到总体成员的概率未知的过程获得的数据进行拟合。野生动物调查数据的一个显著特点是，它们恰恰涉及这种未知的检测概率，更糟糕的是，它们涉及的检测概率在空间上各不相同。为了对种群的空间分布做出可靠的推断，必须将感兴趣种群(S)的空间分布(S)和探测概率的空间分布分开。距离抽样(DS)和捕获-重新捕获(CR)方法是最广泛使用的野生动物调查方法。DS的大部分研究工作都集中在开发可靠的空间检测概率估计方法上。直到最近，CR方法完全忽略了检测概率的空间分量，但随着最近空间显式捕获-再捕获(SECR)方法的出现，CR方法现在也能够估计空间检测概率。但是(除了少数例外)，这两种方法目前都是在假设不现实的简单人口空间分布的情况下估计发现概率的。虽然对丰度的估计对此是可靠的，但对分布的估计则不是。这个项目结合了DS和CR方法以及INLA的优点。它将统一INLA中的空间建模方法以及SECR和DS方法的空间探测概率估计方法，首次提供严格的统计方法和软件，用于使用两种最广泛使用的野生动物调查方法的数据来估计实际复杂的空间分布。它将提供比目前可用的更强大的方法和工具来推断是什么驱动了动植物分布和分布的变化。通过这样做，它将为监测和管理生物多样性丧失提供比目前可用的更强大的工具。由于DS和CR调查通常记录空间数据，这些方法将追溯适用于许多现有的调查数据的时间序列，因此它们可以立即用于“回顾过去”，并与可靠的数据集一样，对追溯到过去的分布和分布变化做出推断。

A reduction of biodiversity loss is a key aim of the Convention on Biological Diversity (CBD) for 2020, and quantifying the loss is essential for managing it. This involves estimating the size and distribution of wild populations, which is statistically challenging - using only animals detected (often a very small fraction of the population), one must deduce the abundance and distribution of animals that were not detected.Natural systems invariably have spatial structure, and monitoring and understanding what drives habitat use, spatial distribution and changes in spatial distribution is central to understanding and predicting the effects of natural or human-induced perturbations of natural systems. This is difficult because the spatial structure of fauna and flora is often complex, involving spatial trend, spatial randomness and spatial correlation. Fitting spatial models that cannot accommodate all these aspects of spatial distribution can lead to very misleading conclusions about the drivers of spatial distribution and changes in distribution. In particular, inadequate modelling of randomness and correlation can lead to incorrect inferences and misleading predictions. And while realistically complex spatial models have existed for some time, until very recently the methods for fitting such models were too slow to be useful. With the advent of the Integrated Nested Laplace Approximation (INLA) method this is no longer the case, and as a result, use of this method has grown rapidly and the software implementing it is in great demand.However, there are currently no methods or software (INLA or other) for fitting realistically complex spatial models to data obtained from processes in which the probability of detecting population members is unknown. And a distinguishing feature of wildlife survey data is that they involve exactly such unknown detection probabilities, and what is worse, they involve detection probabilities that vary in space. The spatial distribution(s) of the population(s) of interests and the spatial distribution of detection probability have to be separated in order to draw reliable inferences about the population spatial distribution.Distance sampling (DS) and capture-recapture (CR) methods are far and away the most widely-used wildlife survey methods. Much of DS research effort has focused on developing methods for reliable estimation of spatial detection probability. Until very recently CR methods neglected the spatial component of detection probability entirely, but with the recent advent of Spatially Explicit Capture-Recapture (SECR) methods, CR methods are now also able to estimate spatial detection probability. But (with a few exceptions) both methods currently estimate detection probability assuming unrealistically simple population spatial distributions. While estimates of abundance are robust to this, estimates of distribution are not.This project combines the strengths of DS and CR methods and INLA. It will unite spatial modelling methods in INLA and spatial detection probability estimation methods of SECR and DS methods, to provide for the first time rigorous statistical methods and software for estimating realistically complex spatial distributions using data from the two most widely-used wildlife survey methods. It will provide more powerful methods and tools than are currently available for drawing inferences about what drives the distribution and change in distribution of fauna and flora. In so doing, it will provide substantially more powerful tools for monitoring and managing biodiversity loss than are currently available. And because DS and CR surveys usually record spatial data, the methods will be retrospectively applicable to many existing time series of survey data, so that they can be used immediately to "look into the past" and draw inferences about distribution and changes in distribution stretching as far back into the past as do reliable data sets.