Novel Machine Learning Methods for Estimating the Spatial Distribution of Species

我的项目重点是开发新的机器学习方法来预测物种的地理分布。这些方法将在大规模的众包数据集上进行训练，并将重点放在我们拥有最少可用数据的物种上。由此产生的产出将使人们更好地了解以前未绘制的物种的分布范围。提高我们对稀有物种分布的了解可能有助于保护工作，因为这有助于确定对物种生存重要的区域，并可以为保护区的设计提供信息。开发的模型还将用于修改来自图像分类器的预测，以帮助消除主要出现在不同地区（例如世界不同地区）的物种之间的歧义。我提出的研究将利用深度学习的最新进展，并允许在同一模型中捕获不同物种的信息。与现有的方法相比，这种方法有几个优点，比如更好地预测更稀有的物种，更好的时间建模，以及利用这些模型确定未来数据收集的最佳采样地点的能力。在这个项目的过程中，将开发的新方法包括创新的方法，以充分利用可用于这项任务的嘈杂和空间偏差数据，模拟物种分布时间变化的数据有效方法，以及在我们目前对物种分布的理解的指导下，为我们目前知之甚少的物种收集有用的额外数据的新方法。一个主要目标是利用最近的深度学习进展来创建新的模型，这些模型可以同时代表物种分布的空间和时间趋势，使我们能够告诉物种目前存在于哪个区域，以及我们预计未来会如何变化，这可能有助于保护应用。另一个目标是开发新的机器学习方法，以最好地利用为这项任务通常收集的不寻常的“仅存在”数据，其中只有观察到物种的位置被记录，而物种不存在的“缺席”必须从存在数据中推断出来。

My project focuses on the development of novel machine learning methods for predicting the geographic distribution of species. These methods will be trained on large-scale, crowdsourced datasets, and will focus on species for which we have the least data available. The resulting outputs will provide a greater understanding about the ranges of previously unmapped species. Improving our knowledge of the distribution of rare species may help with conservation efforts, as this helps to identify areas that are important for species survival and can inform the design of protected areas. The developed models will also be used to modify predictions from image classifiers to help disambiguate between species that mostly appear in different regions (e.g. different parts of the world). My proposed research will avail of recent advances in deep learning and will allow information about different species to be captured inside of the same model. This has several advantages over existing methods such as better predictions for rarer species, better temporal modelling, and the ability to use these models to determine the best sites to sample for future data collection.Novel methodology that will be developed in the course of this project include innovative ways to make best use of the noisy and spatially biased data available for this task, data efficient methods of modelling temporal changes in the distribution of species, and new methods to gather useful additional data for species that we know little about currently, guided by our present understanding of the species distribution. One main objective is using recent deep learning advances to create new models that can simultaneously represent both spatial and temporal trends in the distribution of a species, allowing us to tell both which regions a species is currently present in, and how we expect that to change in the future, which may assist in conservation applications. Another objective is to develop new machine learning methods to best make use of the unusual "presence only" data commonly collected for this task, where only locations that a species has been observed are recorded and "absences" where the species is not present must be inferred from the presence data.