ML-Based Land Use and Land Cover Classification: Assessing Performance and Predicting Future Changes

Abstract

Land use and land cover (LULC) changes can exacerbate the effects of climate change by altering local ecosystems, hydrological cycles, and carbon sequestration processes, particularly in arid and semiarid regions. In the present study, LULC was classified by employing Google Earth Engine (GEE)-based five machine-learning (ML) models, i.e., support vector machine (SVM), classification and regression trees (CART), naïve bayes (NB), gradient tree boost (GTB), and random forest (RF) on the Landsat 8 satellite imagery spanning from 2015 to 2022 for the Rel River watershed. The ML models were validated employing the figure of merit technique, and the accuracy was assessed using the Kappa coefficient and correlation matrix. The results indicate that the RF algorithm has the best accuracy at 95% (K=0.94), followed by GTB at 89% (K=0.86). Further, LULC was predicted for the year 2027 by implementing an artificial neural network (ANN) tool available in the QGIS MOLUSCE plugin. The statistical value of validation methods in terms of receiver operating characteristics (ROC)/area under the curve (AUC) (0.7) and chi-square results reveal that ANN predicted LULC changes appropriately. High AUC-ROC and appropriate chi-square values indicate that the ANN’s predictions were both accurate and statistically robust, supporting its potential for reliable LULC forecasting up to 2027. Conclusively, the RF was the best ML classifier among the five methods; however, it will need to be tested in many future morphoclimatic scenarios. The study highlights the importance of advanced geospatial technologies, ML, and GEE in land use planning and environmental management. This work can assist policymakers in urban planning by providing accurate, up-to-date maps to guide infrastructure development and land zoning decisions.

Topics & Concepts