Predicting land-use change: Intercomparison of different hybrid machine learning models

Abstract

Land Use Land Cover (LULC) change assessment and prediction are essential for optimised water resources planning and management. This paper attempts to intercompare the different LULC change modelling techniques (two-hybrid and two traditional models) to predict the future LULC change. These include Multi-Layer Perceptron Markov Chain (MLP_MC), Logistic Regression-Markov Model (LR-MC), and two hybrid models, namely, Multi-Layer Perceptron Markov Chain Cellular Automata (MLP_MC_CA), Logistic Regression Markov Model Cellular Automata (LR_MC_CA) models. These models were tested on Nagavali River Basin (NRB), a river basin in Southern India, which has seen significant land-use changes over the past two decades. Over the past two decades, the study region experienced dominant changes on the order of 17.42% and 15.22% decrease in scrubland and forest, respectively. At the same time, the agricultural land cover is increased by 35.28%. For the LULC prediction, the model was initially trained using the relevant driver variables and LULC maps of 2010 and 2015. The calibrated model was validated using the 2020 LULC map. The statistical results in terms of Kappa values and chi-square results reveal that the hybrid model MLP-MC-CA has a better agreement (Kappa coefficient 0.902) compared to the other models. Further, it is also observed that the CA-based models have a better ability to capture spatial connections. After combining the MLP_MC model with the Cellular Automata, the former model was improved by 10.8% in terms of the overall Kappa coefficient. The best model for LULC prediction over the next decade (LULC map of 2030) showed that the forest area would decrease by 9.02%, and the agricultural land would increase by 8.74%. Further, the results from the study indicate that the hybrid machine learning models provide a promising alternative for land-use change prediction.