Prediction through Description
Deep Generative Geographical Process Models
Making the case for deep generative geographical process models
Modeling the distributions of geographical processes
Deep generative geographical process models aim to learn the underlying rules which determine geographical processes as diverse as changing weather conditions or traffic patterns directly from spatio-temporal sensor data. Deep here refers to Deep Learning, i.e., employing neural networks composed of many layers. Generative models aim to produce a set of consistent, possible settings of random variables in a graphical model given certain observations. Specifically, this can be employed to generate realistic samples of a complex, real world distribution having only observed samples of their distribution of data as diverse as images, text, music, or videos.
If the sampled2 data originates from a true data distribution p(x), a generative model aims to approximate it with a generated distribution p̂ϑ(x) based on model parameters ϑ. Particularly promising representatives of this approach are Generative Adversarial Networks (GAN), where a Generator Network (‘forger’) produces new samples from the underlying distribution of the observed data and competes against a Discriminator Network (‘policeman’) which aims to tell apart generated (‘fake’) and observed (‘real’) samples.
Fig. 1: Generative Adversarial Networks (GAN), concept
Humans constantly interact with the world via the anthroposphere – the part of the environment which is made or modified by humans – but also react to ubiquitous natural processes such as day/night cycles, changing weather conditions, or extreme events such as wildfires. Both can be described by geographical processes. A geographical process refers to the dynamic character (over time) of our geographical (spatial) environment. The entities it comprises are thus susceptible to change with regards to their location (i.e., movement), shape (e.g., expansion or contraction), or semantic attributes.
The capability to model and simulate these processes – both natural and human-made – is critical for making rational and sustainable decisions. The field of Geosimulation addresses this task using the traditional modeling process of developing a conceptual model, translating it into a computational model (using paradigms such as Agent-based Models or Cellular Automata), and validating the simulation results with real-world observations. Especially the first step, however, is commonly non-trivial due to the complexity and stochasticity of real-world processes. Furthermore, for numerous applications, deriving fully explanatory models might not even be necessary. Therefore, applying generative modeling paradigms to learn the underlying distributions of dynamic geographical phenomena (i.e., their guiding ‘rules’) without the requirement for any fundamental theory and prior knowledge engineering would be highly desirable.
A Deep Generative Geographical Process Model constitutes an unsupervised learning algorithm which aims to approximate the distribution of geographical processes directly from spatio-temporal sensor data.