Abstract:
This study develops and validates an advanced Cellular Automata-Artificial Neural Network (CA-ANN) model to predict future land use and land cover (LULC) changes in Uganda's ecologically sensitive Murchison Bay Catchment. Leveraging high-resolution (10m) Sentinel2 satellite imagery from 2016-2024, the research first characterizes historical LULC dynamics, revealing a 12% expansion of built-up areas (28.13 km² to 31.44 km²) and alarming 58% wetland loss during the study period. The model incorporates six key spatial drivers - elevation, slope, aspect, population density, and proximity to roads and urban centers - identified through rigorous correlation analysis (VIF < 5) and validated against actual 2024 data with 93.2% accuracy (Kappa = 0.93).
The CA-ANN architecture employs a Multi-Layer Perceptron (MLP) with 15,000 training
samples, 5 hidden layers, and 0.001 learning rate, significantly outperforming traditional
CAMarkov models used in similar Ugandan catchments (Kappa improvement of 0.08-0.12). Projections to 2032 indicate continued urbanization (31.9 km² built-up area) but also identify potential wetland recovery zones under conservation scenarios. These findings provide critical insights for Kampala's Physical Development Plan, particularly regarding: (1) northern subcatchment protection priorities, (2) infrastructure planning to minimize wetland encroachment, and (3) climate-resilient urban design. Methodologically, the study demonstrates the superiority of Sentinel-2 data over Landsat for detecting small-scale changes in tropical urban environments, while substantively contributing to Uganda's Vision 2040 and SDG 15 targets through science-based policy recommendations. The replicable framework addresses key gaps in African LULC modeling by: (1) integrating high-resolution temporal data, (2) quantifying non-linear driver interactions through ANN, and (3) providing validated short-term (2028) and medium-term (2032) projections for adaptive management. Future research directions include incorporating land tenure data and testing hybrid LSTM-CA models for enhanced temporal sensitivity.
