The possibilities of using deep learning for land change modeling
By Ana Paula Aguiar, Christian Requena, Chandrakant Singh, Ingo Fetzer
From December 2 to 6 the course “Land Change Modelling: concepts and practice” took place at the Stockholm Resilience Centre. The participants were PhD and MSc students and researchers. The goals of the course were to expose the students to the different types of Land Change Models (LCM) and their applicability for different goals. The first and last days of the course consisted of open lectures providing an overview on the spectrum of existing modeling approaches and typologies. During the other three days, the students had a mix of lectures and practical exercises using the LuccME/TerraME modeling framework.
As part of the open lectures, at the last day, the invited speaker Christian Requena Mesa, PhD candidate at Department for Computer Science and German Aerospace Center (DLR), University of Jena, Germany provided an introductory background about deep learning approaches and their potential application to land change processes. In his thesis he applies the techniques to Earth Systems Science, specifically to landscape prediction. Following his presentation, we engaged in a discussion about how deep learning techniques could improve existing land change modeling approaches. Here we provide a synthesis of this discussion and emerging ideas.
What is deep learning?
Deep learning evolved from the Machine Learning techniques based on supervised, semi-supervised as well as unsupervised learning methods. In a nutshell, the difference is that with Deep Learning it is not necessary to organize/separate the potential input explanatory variables of the pattern you want your function to predict. Simple machine learning techniques, such as single layer neural network (perceptron architecture), works like a linear regression, in which the weight of each explanatory variable is estimated to best fit the desired output. Neural Networks evolved to include multiple hidden layers (Multiple Perceptron architecture) to capture possible interactions among the input variables. With novel computer architectures, deep learning techniques are currently evolving quite fast, boosted also by the big tech companies and increasing computer power, from the multi-layer perceptron architectures to Convolution Neural Networks (CNN) and Recurrent Neural Networks (RNN).
Potential of Deep Learning in Land Change Modeling
Now with complex human interactions constantly perturbing the natural cycle of the environment in the Anthropocene, it has now become exceedingly difficult to select the features responsible for these perturbations. Complex modeling techniques such as ‘deep learning’ and ‘agent-based modeling’ have now gained an edge over the commonly sought out linear and non-linear parametric techniques (Multiple regression, Generalized additive mixed models, etc.). These techniques are now able to chart-out complex interactions of the system and provide reliable projections. The unsupervised learning mechanism of deep learning techniques have an unparalleled advantage over agent-based modelling in terms of analyzing the non-linear relationship from the data itself and thus greatly reducing the effort to code these interactions.
Machine Learning techniques have long been used models following this generic structure to estimate the transition potential of each cell. For example, Multilayer Perceptron is one of the techniques used in the LCM modules of commercial products such as IDRISI and ArcGIS. More broadly, if multivariate statistical analysis, such as linear and logistic regressions are considered Machine Learning, we can then say it is being used for estimating the potential of change for each cell in most existing cellular models.
Therefore, the first possible application of deep learning would be in the development of innovative methods to estimate the potential (suitability) of each cell for land change transitions, combining it to the existing demand and allocation components which deal with quantity of change and competition among classes. Land change processes are complex in time and space, due to the heterogeneity of actors and social-ecological contexts influencing their decisions. See for example, the images of the Brazilian Amazon at different scales (Figure 2). Estimating the potential is key to the performance of LCM. Capturing the heterogeneity of patterns and processes is a challenge through multivariate statistical techniques is always challenging.
Figure 2 – Example of the heterogeneity of spatial patterns of deforestation in the Brazilian Amazon. These patterns are the result of dynamic interactions within the socio-ecological system in different historical contexts.
In a next step, we could think of applications of deep learning to capture both the temporal and spatial dynamic of land change (without the artificial division in three components as adopted in the current cellular models), maybe linking the learning to different agent responses through time. Certainly, an exciting scientific frontier, given the importance and complexity of adequately representing these processes in models at multiple scales.