A Reconstruction–Segmentation Framework for Robust Tree Cover Mapping in North Korea Using Time-Series Reconstruction Autoencoders

Forests are a critical component of global carbon sequestration, biodiversity, and ecosystem services, making accurate mapping essential for long-term monitoring. In North Korea, limited field access, rugged topography, and inconsistent national statistics necessitate reliable remote sensing–based observation. However, frequent cloud contamination challenges the use of optical time-series imagery for forest monitoring. This study introduces a framework that integrates a ConvLSTM-based autoencoder into a U-Net segmentation model to improve tree cover classification from Sentinel-2 time-series data. The autoencoder was pretrained to reconstruct cloud-contaminated or missing observations using multi-octave Perlin-noise perturbations, providing standardized inputs that enhanced segmentation robustness under noisy conditions. Results show that tree cover accuracy exceeded 96% when all five time steps were available and remained stable (94–95%) even with one missing step. Accuracy declined below 90% with three missing steps but remained above 80%, enabling draft classifications under limited data. Confidence analysis further indicated that model certainty is a practical quality-control metric. Annual mapping for 2019–2024 showed a general increase in tree cover, aligning with reported afforestation efforts in North Korea. Taken together, the framework advances long-term monitoring, carbon accounting, and risk assessment in North Korea, while also enabling robust, region-adapted monitoring in cloud-prone, data-limited settings.