Machine learning analysis of land use change dynamics in Malawi’s Chia Lagoon

Land-use and land-cover (LULC) change is a major driver of ecosystem degradation, particularly in freshwater wetlands where spatially explicit evidence is limited. This study aimed to quantify two decades (2004–2024) of LULC change in Malawi’s Chia Lagoon catchment using machine-learning-based remote sensing approaches. Multi-temporal Landsat imagery (2004, 2014, and 2024) was classified using a Random Forest supervised classification algorithm. LULC changes were assessed using three classes (water, vegetation, and bare land), supported by NDVI and MNDWI analyses. Classification accuracy was evaluated using confusion matrices, overall accuracy, producer’s accuracy, and user’s accuracy, with area estimates adjusted using 95% confidence intervals. Following validation of the final confusion matrices, the Random Forest classifications achieved year-specific overall accuracies of 82.0%, 90.%, and 96.0% for 2004, 2014, and 2024, respectively. Between 2004 and 2024, mapped vegetation decreased from 519 km² (15.4%) to 330 km² (9.8%), representing a net decrease of 189 km² and 5.6% points in catchment share. Bare land increased by 208 km² (+ 6.2% points), whereas mapped water decreased by approximately 20 km² (− 0.6% points). NDVI and MNDWI provided complementary spectral information for interpreting vegetation greenness and surface-water patterns but were not treated as independent validation of the classified area changes. The classified Landsat record indicates significant changes in land-cover composition between 2004 and 2024, including a net reduction in mapped vegetation and a net increase in non-vegetated surfaces, despite a partial increase in vegetation between 2014 and 2024. Because the analysis is based on three dry-season observations and a broad three-class classification scheme, the results should be interpreted as mapped land-cover changes rather than direct evidence of specific causal processes or ecological degradation. Machine-learning-supported remote sensing provides a useful framework for identifying landscape changes that can subsequently be investigated through targeted field-based monitoring.” Not applicable.