Similarity Weighted Instance‐based Learning for the Generation of Transition Potentials in Land Use Change Modeling

Abstract

Land use change models are increasingly being used to evaluate the effect of land change on climate and biodiversity and to generate scenarios of deforestation. Although many methods are available to model land transition potentials, they are usually not user-friendly and require the specification of many parameters, making the task difficult for decision makers not familiar with the tools, as well as making the process difficult to interpret. In this article we propose a simple method for modeling transition potentials. SimWeight is an instance-based learning algorithm based on the logic of the K-Nearest Neighbor algorithm. The method identifies the relevance of each driver variable and predicts the transition potential of locations given known instances of change. A case study was used to demonstrate and validate the method. Comparison of results with the Multi-Layer Perceptron neural network (MLP) suggests that SimWeight performs similarly in its capacity to predict transition potentials, without the need for complex parameters. Another advantage of SimWeight is that it is amenable to parallelization for deployment on a cloud computing platform.t gis_1226 569..580 relationship between land change and driver/suitability variables to evaluate the readiness of the land for transition from one land cover to another (generation of transition potentials), the determination of the amount of land that will transition (demand of land) and finally, the allocation of land to the new land cover types (land allocation). The evaluation of transition potentials is a critical step in the process of land change modeling and prediction, as the final allocation of land is based upon them. Therefore the production of transition potentials with high accuracy is essential in the process of land change prediction. Although many methods exist for the generation of transition potentials, to our knowledge, only Eastman et al. (2005) have evaluated an extensive list of them. In their work they evaluated (among others) the Weights of Evidence, which is used in DINAMICA (Soares-Filho et al. 2002); empirical probabilities, which is used in GEOMOD (Pontius et al. 2001); logistic regression, which is used, for example, in CLUE-S (Verburg et al. 1999) and the Multi-Layer Perceptron neural network, which is used in LTM (Pijanowski et al. 2002) and LCM (Eastman 2009). The main conclusion of that work is that there are large differences in the patterns of transition potentials generated by the different methods, and therefore they are not interchangeable. Moreover, in their tests, they found that the Multi-Layer Perceptron (MLP) performed best.Non-parametric methods such as the Multi-Layer Perceptron are appealing for the generation of transition potentials since they can fit complex relationships, and they are capable of producing accurate results. However, in order to train the neural networks for the determination of transition potentials, many parameters need to be adjusted, which makes the model difficult to run and complex to understand....

Modeling the Impact of Urban Landscape Change on Urban Wetlands Using Similarity Weighted Instance-Based Machine Learning and Markov Model

“…Many studies have integrated Markov chain with other transition potential modeling methods such as the multi-layer perceptron neural network and cellular automata methods (e.g., [25,26]). This study utilized the Similarity Weighted Instance-based Machine Learning approach in combination with Markov chain because it has the capacity to predict transition potentials without the need for complex parameters [24] associated with other methods. The results obtained from integrating both methods of landscape simulation were used in assessing the impact of urban expansion induced landscape change on wetlands within the study area.…”

“…Each variable weight is then determined by comparing the standard deviation of the variable inside areas that have changed to the standard deviation of the variable for the entire study area [24]. A smaller standard deviation of the variable inside the pixels that have changed when compared to the study area as a whole signifies that the variable is relevant for discriminating change [24]. Figure 8 is the relevance weight of each variable that is associated with the six transitions determined for the three watersheds.…”

“…Variable weighting is calculated by multiplying each variable by a weight that is proportional to its ability to distinguish between classes [35]. Each variable weight is then determined by comparing the standard deviation of the variable inside areas that have changed to the standard deviation of the variable for the entire study area [24]. A smaller standard deviation of the variable inside the pixels that have changed when compared to the study area as a whole signifies that the variable is relevant for discriminating change [24].…”

“…The TerraSet Geospatial Monitoring and Modeling System's Land Change Modeler (LCM) is an innovative land planning and decision support software tool that simplifies the complexities of land change analysis and prediction [34]. The SimWeight model, well described in Sangermano et al [24], is an instance-based machine-learning algorithm that is based on a modified K-nearest neighbor machine learning algorithm, which calculates the weighted distance in variable space to known events for the classes. The method identifies the relevance of each land change-driving variable in the change process and predicts the transition potential of locations given known instances of change [24].…”

“…Based on the simulation results, the study assessed the impacts of future landscape change on urban wetlands. The Similarity Weighted Instance-based Machine Learning, described by Sangermano et al [24] was used in predicting the potential for the locations being examined to change from one time to another time using known instances of change in the study area. The outputs from this process are maps representing the potential for change in the locations under examination.…”