Deep Learning Based Monitoring of Forest Dynamics and Degradation in Agroforestry Landscapes of Ethiopia:A Systematic Review.

Agroforestry landscapes in Ethiopia are undergoing rapid transformation driven by agricultural expansion, biomass extraction, and climate variability. Conventional monitoring approaches remain inadequate for capturing fine-scale forest dynamics in heterogeneous smallholder systems. This systematic review applies the PRISMA 2020 framework to evaluate the application of deep learning in monitoring forest dynamics and degradation in Ethiopian agroforestry landscapes. A comprehensive search across Scopus, Web of Science, Google Scholar, NASA ADS, and Semantic Scholar through 15 March 2025 identified fifteen eligible studies published between 2018 and early 2025. Results indicate that attention-based convolutional architectures applied to high-resolution imagery achieve accuracy levels between 85 and 89 percent for post-deforestation land-use classification in Ethiopia. Hyperspectral data from the PRISMA satellite combined with one-dimensional convolutional neural networks achieve over 90 percent accuracy for deforestation and degradation detection in comparable tropical systems. Neural network models using Sentinel-2 time series show moderate performance (80–85 percent) for degradation detection in Ethiopian coffee agroforestry systems. Random Forest remains a strong benchmark for tree species mapping in Eucalyptus woodlots, reaching 96.3 percent overall accuracy. Despite these advances, significant limitations persist, including spectral ambiguity in mixed systems, limited labeled datasets for the Ethiopian highlands, and insufficient validation frameworks. The review concludes that deep learning offers strong potential for operational forest monitoring and REDD+ implementation in Ethiopia but requires localized datasets, multi-sensor integration, and institutional investment in computational infrastructure.