Explainable Population Estimation Using Deep Learning from Satellite Imagery

联合国制定的可持续发展目标（SDG）中，超过三分之一的指标是根据总人口或特定的人口亚群来定义的。一个地区的最新人口信息对于决策至关重要，包括获得服务，分发疫苗，救灾等。传统的人口数据，如人口普查，不足以达到这一目的，因为人口普查通常每十年进行一次，而最需要最新人口统计的国家进行普查的频率更低。近年来，利用卫星图像等替代数据来源进行人口估计受到了极大关注。人口普查依赖[2]和人口普查独立[3]方法都已取得了一定的成功，其中许多方法利用了先进的图像分析方法，如深度卷积神经网络，结果令人鼓舞。该项目将进一步开发这些方法，以产生可持续的，可解释的和可靠的机器学习模型，更有效地估计人口。这将涉及利用空间和时间背景信息，将多种分辨率的卫星图像与土地覆盖图等公开数据源相结合，该项目的目的是开发可持续的、可解释的和可靠的机器学习模型，以便利用卫星图像和调查信息有效地估计一个地区的人口。该项目将调查以下研究问题：（1）上下文邻里信息，无论是短期和长期的，可以通过了解周围地区的特点，改善人口估计？(2)人口普查、调查和微型人口普查的信息能否结合起来，以可靠地跟踪人口的变化？(3)能否将来自不同来源和不同分辨率的数据结合起来，以获得关于一个地区的补充信息？(4)与人口稠密的城市地区相比，人口稀少的农村地区的不确定性/偏见是否有所不同？以及（5）估计的人口和相关的不确定性能否有效地向政策制定者解释？该项目将利用计算机视觉模型、深度学习架构和可解释的模型构造，结合不同分辨率的卫星图像和调查数据，以估计一个地区的人口。

More than one-third of the Sustainable Development Goals (SDGs) indicators established by the United Nations (UN) are defined in terms of total population or a specific demographic sub-population [1]. Up-to-date population information of a region is crucial for decision making including access to services, distribution of vaccinations, disaster relief, and many others. Traditional population data, such as census, are not adequate for this purpose since censuses are typically conducted decennially and countries with the greatest need for up-to-date population counts conduct them even less frequently. Population estimation using alternative data sources such as satellite imagery has received significant attention in the recent years. Both census-dependent [2] and census-independent [3] approaches have been explored with some success, and many of these methods have utilized advanced image analysis methods such as deep convolutional neural networks with promising results. This project will develop these methods further, to produce sustainable, interpretable and reliable machine learning models estimating population more effectively. It will involve utilizing contextual information, both spatial and temporal, integrating satellite imagery at multiple resolutions with publicly available data sources such as land cover map, etc., combining from census, surveys and microcensus data, and explaining the decisions made by these models to the end-users.The aim of the project is to develop sustainable, interpretable, and reliable machine learning models to effectively estimate the population of an area using satellite imagery and survey information. The project will investigate the following research questions: (1) Can contextual neighbourhood information, both short-range and long-range, improve population estimates by understanding the characteristics of the surrounding regions? (2) Can information from census, surveys, and micro-census be combined to track population reliably over time? (3) Can data from different sources and different resolutions be combined to acquire complementary information about an area? (4) Does uncertainty/bias differ in sparsely populated rural areas vary compared to densely populated urban areas? and (5) Can estimated population and associated uncertainty be explained to policymakers effectively?MethodologyThe project will incorporate satellite imagery of different resolutions and survey data using computer vision models, deep learning architecture and explainable model constructs to estimate population of a region.