Comparative evaluation of geospatial scenario-based land change simulation models using landscape metrics

Abstract

Assessing the performance of land change simulation models is a critical step when predicting the future landscape scenario. The study was conducted in the district of Varanasi, Uttar Pradesh, India because the city being “the oldest living city in the world” attracts a vast population to reside here for short and long-term, leaving the city’s ecosystem more exposed to fragility and less resilient. In this work, an approach based on landscape metrics is introduced for comparing the performance of the ensemble models designed to simulate the landscape changes. A set of landscape metrics were applied in this study that offered comprehensive information on the performance of scenario-based simulation models from the viewpoint of the spatial ordering of simulated results against the related reference maps. A supervised support vector machine classification technique was applied to derive the LULC maps using Landsat satellite images of the year 1988, 2001, and 2015. The LULC maps of 1988 and 2001 were used to simulate the LULC scenario for 2015 using three Markov chain-based simulation models namely, multi-layer perceptron-Markov chain (MLP_Markov), cellular automata-Markov chain (CA_Markov), and stochastic-Markov chain (ST_Markov) respectively. The mean relative error (MRE), as a measure of the success of simulation models, was calculated for metrics. The MRE values at both the class and landscape levels were accounted for 21.63 and 11.45% respectively using MLP_Markov simulation model. The MRE values at both the class and landscape levels were accounted for 39.61 and 28.31% respectively using CA_Markov simulation model. The MRE values at both the class and landscape levels were accounted for 55.36 and 45.75% respectively using ST_Markov simulation model. The MRE values considered at class and landscape levels are further evaluated qualitatively for comparing the performance of simulation models. The results indicate that the MLP_Markov performed excellently, followed by CA_Markov and ST_Markov simulation models. This work showed an ordered and multi-level spatial evaluation of the models’ performance into the decision-making process of selecting the optimum approach among them. Landscape metrics as a vital characteristic of the utilized method, employ the maximum potential of the reference and simulated layers for a performance evaluation process. It extends the insight into the main strengths and drawbacks of a specific model when simulating the spatio-temporal pattern. The quantified information of transition among landscape categories also provides land policy managers a better perception to build a sustainable city master plan.