Multi-Sensor Mapping of Forest Type Transitions in Central Angola: A Cloud-Based Random Forest Assessment (2015–2025)

Miombo woodlands are critical for biodiversity conservation and climate regulation in Central and Southern Africa, yet they are increasingly threatened by agricultural expansion, logging, and fire-related disturbances. In Central Angola, limited spectral separability between natural Miombo formations and timber plantations constrains accurate monitoring of forest transitions. This study developed a cloud-based multi-sensor machine learning framework to assess land use/land cover and forest-type dynamics between 2015 and 2025, contributing to SDG 13 and SDG 15. Landsat 8 (2015–2016) and Sentinel-2 (2017–2025) imagery were processed in Google Earth Engine and Google Colab Pro to generate annual cloud-free median composites (p50). Key surface reflectance bands and spectral vegetation and disturbance indices (NDVI, EVI, NDWI, NBR, and NDTI) were integrated into a Random Forest classification workflow. Training samples were derived from K-means clustering and refined through historical high-resolution visual interpretation using a stratified 70/30 train–validation split. Classification performance was evaluated using Overall Accuracy (OA), Kappa coefficient, and Quantity and Allocation Disagreement metrics. Results revealed pronounced forest loss and landscape transformation across the decade. Open Miombo declined from 42,330 km² in 2015 to 30,719 km² in 2025, corresponding to a significant loss rate of −1,007 km² year⁻¹ (Sen’s slope; p = 0.016) and an overall reduction of 27.4%. Miombo Wooded Savannas decreased by approximately 40%, while Grasslands expanded by 166.4%, indicating a progressive transition from forested formations to open vegetation systems. Smallholder agriculture increased by 19% and showed a negative correlation with Open Miombo extent (r = −0.40), confirming agricultural expansion as a major driver of forest conversion. Timber plantations displayed the highest temporal variability (CV = 61.5%), reflecting cyclical harvesting and replanting dynamics, with plantation extent recovering from 34 km² in 2023 to 135 km² in 2025. Classification accuracy remained robust throughout the study period (OA: 0.68–0.91; Kappa: 0.63–0.90).