Performance analysis of radial basis function networks and multi-layer perceptron networks in modeling urban change: a case study

The majority of cities are rapidly growing. This makes the monitoring and modeling of urban change’s spatial patterns critical to urban planners, decision makers, and environment protection activists. Although a wide range of methods exists for modeling and simulating urban growth, machine learning (ML) techniques have received less attention despite their potential for producing highly accurate predictions of future urban extents. The aim of this study is to investigate two ML techniques, namely radial basis function network (RBFN) and multi-layer perceptron (MLP) networks, for modeling urban change. By predicting urban change for 2010, the models’ performance is evaluated by comparing results with a reference map and by using a set of pertinent statistical measures, such as average spatial distance deviation and figure of merit. The application of these techniques employs the case study area of Mumbai, India. The results show that both models, which were tested using the same explanatory variables, produced promising results in terms of predicting the size and extent of future urban areas. Although a close match between RBFN and MLP is observed, RBFN demonstrates higher spatial accuracy of prediction. Accordingly, RBFN was utilized to simulate urban change for 2020 and 2030. Overall, the study provides evidence that RBFN is a robust and efficient ML technique and can therefore be recommended for land use change modeling.     

Artificial neural networks and cluster analysis in landslide susceptibility zonation

A landslide susceptibility analysis is performed by means of Artificial Neural Network (ANN) and Cluster Analysis (CA). This kind of analysis is aimed at using ANNs to model the complex non linear relationships between mass movements and conditioning factors for susceptibility zonation, in order to identify unstable areas. The proposed method adopts CA to improve the selection of training, validation, and test records from data, managed within a Geographic Information System (GIS). In particular, we introduce a domain-specific distance measure in cluster formation. Clustering is used in data pre-processing to select non landslide records and is performed on the whole dataset, excluding the test set landslides. Susceptibility analysis is carried out by means of ANNs on the so-generated data and compared with the common strategy to select random non-landslide samples from pixels without landslides. The proposed method has been applied in the Brembilla Municipality, a landslide-prone area in the Southern Alps, Italy. The results show significant differences between the two sampling methods: the classification of the test set, previously separated and excluded from the training data, is always better when the non-landslide patterns are obtained using the proposed cluster sampling. The case study validates that, by means of a domain-specific distance measure in cluster formation, it is possible to introduce expert knowledge into the black-box modelling method, implemented by ANNs, to improve the predictive capability and the robustness of the models obtained.