<![CDATA[Global deforestation and land-use change require timely, accurate, and transferable monitoring systems capable of processing extensive satellite archives. This study aimed to evaluate deep learning architectures for automated classification of satellite imagery, detection of forest change, and identification of post-deforestation land uses across diverse ecological regions. A comparative experimental design was applied to 120,000 labelled image patches derived from Landsat, Sentinel-1, and Sentinel-2 data representing South America, Southeast Asia, Central Africa, North America, Europe, and Oceania. Nine architectures, including convolutional, recurrent, Siamese, Transformer-based, and hybrid models, were assessed using spatial, temporal, and cross-regional validation. The hybrid CNN–Transformer achieved the strongest overall accuracy, calibration, and transfer performance, while radar–optical fusion improved classification in persistently cloudy tropical environments. Multitemporal models reduced confusion between permanent deforestation, seasonal disturbance, harvesting, and regeneration. Geographic holdout testing produced substantial performance declines across all architectures, demonstrating that random partitioning overestimated operational reliability. The study concludes that effective global deforestation monitoring requires multimodal data, temporal context, uncertainty estimation, active learning, and geographically independent evaluation. Automated classification can strengthen forest surveillance, yet expert verification and statistically adjusted area estimation remain essential for responsible environmental decision-making and policy implementation under rapidly changing climatic, technological, and land-management conditions worldwide over time.]]>
Copyrights © 2026
