Advanced characterization of deforestation frontiers in Nigeria utilizing deep learning and Bayesian approaches with sentinel-1 SAR imagery

The rapid deforestation in Nigeria poses significant environmental challenges, including biodiversity loss and degradation of vital ecosystem services. While previous studies have identified potential causes of deforestation, accurately tracking fine-scale temporal patterns and mapping degradation remain difficult. Limitations include infrequent high-resolution satellite observations, obscured understory disturbances, and the shortcomings of conventional reflectance-based remote sensing methods, which typically focus on annual changes. Addressing these challenges, we propose an innovative approach that integrates spatial and temporal features for biannual deforestation mapping using synthetic aperture radar data in hotspots, study site 1 (Akure) and study site 2 (Okomu) forest reserve (First and second halves of 2020, 2021, 2022, and 2023). Our method includes: (1) evaluating a spatiotemporal deep learning algorithm, particularly U-Net and DeepLab3, with various ResNet backbones across eight temporal datasets for delineating deforestation; (2) employing a Bayesian temporal advancing approach to combine deforestation detection from the Sentinel-1 and (3) conducting frontier analysis based on biannual results to assess patch size distribution and formation speed from 2020 to 2023. The findings indicate that the U-Net with ResNet34 outperformed other deep learning models, achieving a precision of 86.5%, recall of 94.5%, an F1 score of 91.3%, and an IoU of 74.2%. The Bayesian approach improved mapping accuracy by reducing uncertainties in post-deforestation observations and the deforestation temporal clustering indicates disguisable between the study sites: deforestations study site 2 (Okomu) has a higher concentration during the second half of 2023 Coefficient Variation (CV= 1.939), Study site 1 (Akure) Coefficient Variation (CV= 1.786), and temporal distribution in deforestation varies within each year. These methods provide critical temporal insights into deforestation dynamics, essential for understanding ecological impacts and developing effective conservation strategies.