Machine learning (ML) entered probably all disciplines which use quantitative methods. Its progress is often driven by the needs of specific data and questions that should be answered. Regional science and spatial data are also experiencing the inflow of machine learning methods into its toolbox. However, the spatial dimension requires exceptional treatment and redesigning the methods when transferred from other areas. Especially, that the new sources as OpenStreetMap and GoogleMaps with background maps, points of interest, roads, traffic, etc., as well as geo-referenced images as satellite photo, night light photo, drone photo, and also geotagged social media posts on Twitter or climatic sensors, are very demanding.
This paper is a methodological overview of spatial machine learning. It catalogues and summarizes the existing solutions which treat spatial data with machine learning tools. It analyses the algorithms and spatial “tricks” used together with typical ML methods as well as the new methods designed to address spatial challenges. The paper goes through unsupervised learning with clustering methods, and supervised learning with classification and regression models. The goal of the paper is to put the line between problems solved and still waiting for solutions. It is the methodological study which balances between classical and spatial statistics, econometrics and machine learning.
There are three essential aspects of this research. Firstly, these are the scientific implementations of spatial machine learning studies and outlining what kind of new questions one can ask when using ML methods in regional science. Secondly, it shows a set of old methods that combined in a new framework of spatial ML work fine. Paper proves that spatial ML is like LEGO construction in many applications, consisting mostly of well-known small parts and occasionally with new concepts added. Third, it is to talk about available software implementations, which make the computations feasible.
Paper is based on the functional approach. The methods are analysed not due to their formal classification but the general mechanism, input and output. This pragmatic way of presentation seems to be attractive for spatial non-methodologists, who are simply curious about what else can be done with data and what knowledge can be found.