COMPARATIVE ASSESSMENT OF RANDOM FOREST AND SUPPORT VECTOR MACHINE FOR LULC CLASSIFICATION AND CNN AND CA-MARKOV FOR PREDICTION OF LULC DYNAMICS IN DEDZA BOMA, MALAW

Land Use and Land Cover (LULC) change is a major environmental concern in rapidly developing towns such as Dedza Boma, Malawi. Understanding how land cover has changed over time is essential for informing sustainable urban planning and natural resource management. This study conducted a comparative assessment of machine learning approaches for LULC classification and prediction in Dedza Boma for the years 2015, 2020, and 2025. Sentinel-2 satellite imagery was classified using two algorithms Random Forest (RF) and Support Vector Machine (SVM) on the Google Earth Engine (GEE) platform. Accuracy assessment revealed that RF outperformed SVM across all time periods, achieving overall accuracies of 90.0% (2015), 91.0% (2020), and 92.0% (2025) with Kappa coefficients of 0.87, 0.88, and 0.89 respectively, compared to SVM accuracies of 86.0% (2015), 87.0% (2020), and 88.0% (2025). RF was therefore selected as the best-performing classification algorithm.
Post-classification change detection revealed notable land cover transitions over the study period. Major changes included the conversion of farmland and vegetated areas into built-up land, indicating rapid urban expansion within the town. Additional shifts from vegetation to bare land and farmland were also observed, demonstrating increasing pressure on natural resources. For future LULC prediction, two approaches were comparatively assessed: a Convolutional Neural Network (CNN) deep learning model and the Cellular Automata–Markov (CA–Markov) model. CNN outperformed CA–Markov, achieving an overall accuracy of 89.0% (Kappa: 0.87) compared to CA– Markov’s 85.0% (Kappa: 0.82). Future LULC scenarios produced using CNN showed continued growth of built-up areas and further declines in farmland and vegetation by 2030 and 2035.
The overall findings provide up-to-date and reliable spatial information that can guide planners, policymakers, and environmental managers in making informed decisions. The study concludes that RF is the best-performing classification algorithm and CNN is the superior prediction model for LULC dynamics in Dedza Boma. The predicted LULC scenarios offer valuable insights for sustainable land use planning and future development in the study area.

This study conducted a comparative assessment of machine learning approaches for LULC classification and prediction in Dedza Boma for the years 2015, 2020, and 2025. Sentinel-2 satellite imagery was classified using two algorithms Random Forest (RF) and Support Vector Machine (SVM) on the Google Earth Engine (GEE) platform. Accuracy assessment revealed that RF outperformed SVM across all time periods, achieving overall accuracies of 90.0% (2015), 91.0% (2020), and 92.0% (2025) with Kappa coefficients of 0.87, 0.88, and 0.89 respectively, compared to SVM accuracies of 86.0% (2015), 87.0% (2020), and 88.0% (2025). For future LULC prediction, two approaches were comparatively assessed: a Convolutional Neural Network (CNN) deep learning model and the Cellular Automata–Markov (CA–Markov) model.

Dedza Boma is the administrative and commercial center of Dedza District in Central Malawi (Munthali et al., 2019). It lies approximately between latitudes 14.3776°S and longitudes 34.3326°E, covering an area of about 45 square kilometers (National Statistical Office, 2018; MASDAP, 2020). The area is characterized by a mixture of urban, peri-urban, agricultural, and semi-natural environments, having undergone rapid urbanization and infrastructure development in recent years leading to significant land cover transformations (Munthali et al., 2019, 2022). These dynamics make it an ideal site for assessing spatio-temporal LULC patterns and their environmental consequences (Lambin & Geist, 2006; Munthali et al., 2022).

Traditional authority: TA Kaphuka

District: Dedza

Coordinates: -14.2306, 34.12585