Several ML algorithms have been tested for their performance on different kinds of datasets for land use classification and simulation of land use planning processes. The more popular algorithms are support vector machine, neural network, Markov random field, GANS and random forest. These algorithms are experimented on different data sets individually and in combination. The article is a review of the functionalities of these algorithms and their application in land use planning.

There is ongoing research for new methods of ML to take land use mapping to a higher plane. Support vector machines (SVMs) have been applied in a number of research papers and have been compared for their performance in land use classification with other ML algorithms such as random forest (RF), neural network. SVMs is a group of non-parametric ML algorithms. The core operation of SVMs is to construct a separating hyperplane (i.e., a decision boundary) on the basis of the properties of the training samples, specifically their distribution in feature space. In many instances, classification in high dimension feature spaces results in over-fitting in the input space, however, in SVMs over-fitting is controlled through the principle of structural risk minimization. The empirical risk of misclassification is minimized by maximizing the margin between the data points and the decision boundary [5]. In case of computational requirements SVMs work well with small data sets with fewer outliers [6]. Among the decision tree algorithms like CART (classification and regression tree), ID3 (iterative dichotomizer 3) more commonly used algorithm for land use classification is RF. One of the benefits of RF algorithm is that it can be used for both classification and regression. RF works as an ensemble learning algorithm based on decision tree classifiers, bagging, and bootstrapping. Each tree is trained by bootstrapping, using different samples from the training data. Additionally, each tree is trained using a random subset of the predicting variables. RF may use thousands of decision trees, where each tree casts a vote and the prediction of the class is decided by the majority vote. A big committee of randomly created decision tree determines the classification, hence the name, random forest [7]. RF can handle large number of variables without need of deleting any and bringing out the relative importance of each of the variables. As compared to SVMs, RF do not have hyperparameters to tune like choosing the right kernel, regularization, penalty, the slack variable, however, the complexity and computational cost increase with the increase in the number of trees in the forest. A list of some of the most commonly used variables is given in Table 2.

Apart from SVM and RF, another ML algorithm which has been widely applied for land use classification is deep learning methods. In 2006, deep learning was introduced by Hinton et al. (2015) [10]. Deep learning methods are representation learning methods composed of multiple layers and each layer computes a new data representation from the representation in the previous layers of artificial neurons creating a hierarchy of data abstractions [10]. Among the group of deep learning methods is convolutional neural network (CNN) composed of convolution and pooling that are concluded by a fully connected neural network layer and a proper activation function, i.e., in models that directly reconstruct an output image prediction, such as U-Net and generative models, the fully connected network and activation function is not needed [11]. In deep learning, artificial neural network (ANN) has been gaining importance in land use planning studies. ANN is a computational ML model based on multilayer perception composed of processing elements forming three kinds of layer (input, hidden, output) which are called perceptrons [12]. Deep learning algorithms work well with relatively large datasets with supporting infrastructure to train them in reasonable time. There has been increasing interest in the Markov random field-based methods for land use classification and land use change as it helps in generating a smooth classification pattern. Markov random field (MRF) is a statistical model based on probability theory which efficiently represents dependency between pixels in a spatial domain. MRF is useful for characterizing spatial-contextual information and has been commonly used for image segmentation, texture analysis, edge detection and image restoration. MRF has been used for linear feature detection with satisfactory results. MRF for modeling spatial context relies on its relationship to Gibbs random field which is a useful way to apply MRF to deal with context [13]. In the paper “Identifying Urban Poverty Using High-Resolution Satellite Imagery and Machine Learning Approaches: Implications for Housing Inequality” [14], six types of image features perimeter, line segment detector (LSD), Hough transform, gray-level co-occurrence matrix GLCM, HoG, and local binary patterns (LBP) were extracted to identify urban poverty in Wuhan, China. The paper utilizes four machine learning regression approaches random forest (RF), Gaussian process regression (GPR), support vector regression (SVR), and neural network (NN) to study whether the features derived are helpful in differentiating urban poverty. It was concluded in the paper that textural features are important in identifying urban poverty in the study area.

In addition to the above-mentioned ML algorithms, there are several simulation models for the purpose of mapping and growth projections of land use. There are mainly two groups according to the key mechanisms to simulate the process of land use change rule-based/process-based models and empirical-statistic models [15].

Cellular Automata (CA)have been defined as discrete spatio-temporal dynamic systems based on local rules. In cellular models, geographic space is represented in the form of a geographic grid, such as the cells in a raster Geographic Information System. They are preferred when model states and the probabilities of transitions among those states are known and stable. They are most suitable for measuring, detecting and predicting change processes such as land use change and urban growth [16].

Cellular automata have capacity to handle temporal dynamics. Cellular automata have the following basic features:

-

States: each cell can take an integer value that corresponds to the current state of that cell. There is a finite set of states.

-

Neighborhood: is a collection of cells that interact with the current one. To perform simulations on a satellite image we normally take the eight surrounding pixels as neighborhood.

-

Transition function (f): takes as input arguments the cell and neighborhood values and returns the new state of the current cell.

The transition function is applied to each cell of the grid across several iterations. Therefore, cellular automata have an evolution process because some cells are changing their states across the different iterations [17]. The most commonly used cellular automata model is the Slope, Land cover, Exclusion, Urban growth, Transport and Hill shade (SLEUTH) model which has been in application for a long time. SLEUTH model has been widely used for simulating urban growth and land use change. SLEUTH is open source and has been developed in C programming language. As described by Berberoğlu et al. (2016) [18].

Statistical modeling methods are widely used for modeling, assessing, qualifying, quantifying and predicting the degree/extent/direction of land use change and growth. An example of a statistical model is the logistic regression model. Logistic regression is a predictive statistical modeling technique which applies multivariate regression to predict future land use based on historical land use changes, their spatial (change) characteristics and other potential drivers [15]. It is easier to model land use change using statistical modeling methods as their calibration is not so computationally intensive compared to rule-based models like cellular automata [19]. In logistic regression, social and economic factors like population density, accessibility to services, distance to commercial and industrial area, mean incomes, etc., can be incorporated in the model. Logistic regression analysis has been one of the most widely used approaches in the past two decades for predictive land use modeling by means of variation of inductive modeling [19].

Agent-based modeling (ABM), which is a forward-looking simulation technique which calculates “agents,” each of which represents an actor and how they interact with their “environment” or the total system. The models represent real and imagined scenarios, which allow for the discovery of potentially emergent issues or phenomena. Such models have increasingly been used to analyze complex issues like land use change. Agents are independent entities which have set goals to achieve. The agents can be countries, landowners, land tenants, citizens, etc. ABM is used to simulate human behavior in cities, for example whereby policy makers, planners or citizens are entities (agents) which interact with the city environment and are capable of making urban planning decisions [16]. In a rule-based approach, their behavior is fixed, meaning that decision-making functions and algorithms remain unchanged (i.e., they always react in the same way when confronted with a particular situation). While agents react to changes in their spatial and social environment, they neither adapt their rules in response nor intelligently learn from previous experiences. ABM can be useful to study the changes in the land use and to evaluate the projections.

The various development environment used for agent-based modeling are Anylogic, Cormas, Cougaar (via OpenMap) Framsticks, Janus (using JaSIM), MASON, Repast, SeSAm, VisualBots, and NetLogo. These environments provide tools to develop an agent-based model and provide a platform to represent model components, control model function, and evaluate and visualize model output [20].

Hybrid approaches have been developed by integrating different ML methods which has resulted in better performance and assessment. An example of a hybrid method is the integration of logistic regression, Markov chain and cellular automata [21] to model urban expansion in the metropolitan area of Tehran, Iran. The results of the simulation were compared with the actual land use map and the result matched 89% between the simulated and the actual. Another example is integration of cellular automata Markov chain (CA-MC) with artificial neural network [22] to enhance the simulation capacity in predicting the changes in land use. The study integrates ANN and CA-MC to incorporate several driving forces (economic, spatial and environmental variables) that impact land use change. The integration and the influence of the driving forces improved the model prediction. Among the hybrid approaches Kamusoko and Gamba (2015) [23] tested random forest–cellular automata to study urban land change in Harare metropolitan province, Zimbabwe. Cellular automata was used to calculate multiple-step transition rates from land use/land cover maps (1984, 2002 and 2008). RF model was used to compute transition potential maps. The study then compared this model with SVM-CA and logistic regression (LR) and CA. The result showed that RF-CA outperformed SVM-CA and LR-CA models. Hybrid Urban Expansion Model (HUEM) was used by Mustafa and Cools (2018) [24] that integrates LR, CA, AB to simulate future urban development in Wallonia, Belgium. The urban expansion is simulated between 1990 and 2000. The calibration results are analyzed by comparing the projection for the 2000 simulated map with the actual 2000 land use map. The HEUM model uses three agent sets, developer agent, farmer agent and planning permission authority agent. The performance of HEUM is compared with other spatial expansion models, i.e., Logit model, CA model, CA-Logit model. The comparison shows that the performance of HEUM model is better than other models in terms of allocation ability.