Hybrid cellular automata-artificial neural network land use change QGIS-MOLUSCE for urban watershed resilience

Abstract

Accurate land-use change (LUC) forecasts are essential for resilience-oriented planning in fast-urbanising watersheds. This study operationalises an open, reproducible cellular automata-artificial neural network (CA-ANN) workflow within the QGIS MOLUSCE plugin to simulate LUC in the Banjir Kanal Timur (BKT) watershed over 2004– 2034. Multitemporal Landsat scenes (2004, 2014, 2024) were classified via maximum likelihood into five land-use classes, achieving overall accuracy (OA) of 90.7–94.9% and Kappa coefficients of 0.852–0.869; independent validation of the CA-ANN model against the 2024 map yielded Kappa = 0.829, indicating excellent agreement. Empirically, settlement expanded from 27.404 km2 in 2004 to 43.158 km2 in 2024 (>78% of the watershed), while forest declined from 22.467 to 5.775 km2 and water bodies from 3.697 to 2.586 km2. Forward simulation to 2034 indicates further settlement growth to 44.583 km2 (80–81%), with forest contracting to 6.710 km2 (12%) and water bodies to 2.776 km2 (5%), signaling increasing imperviousness, reduced ecological buffering, and heightened flood and habitat- fragmentation risk. The hybrid CA-ANN model reproduces characteristic urban clustering and vegetation fragmentation and generates decision-oriented spatial layers, such as transition-potential surfaces and maps highlighting areas of rapid conversion and increased exposure. By embedding CA-ANN in an accessible QGIS- based pipeline, the study advances a transferable decision-support approach that links quantitative LUC forecasts and their uncertainties to enforceable growth boundaries, riparian buffers, and portfolios of low-impact development. Consequently, strengthens evidence-based zoning and urban watershed resilience planning.