National scale sub-meter time series mangrove mapping using Landsat imagery and deep transfer learning

Ma Junkai, Chunyuan Diao, Jinyan Tian, Meng Xu, Mingming Jia, Chen Shi, Jie Song, Zhaorong Liu, Jiahuan Liu, Wei Chen, Songzhao Mei, Xiumin Zhu, Niu Luo, Lin Zhu

Current mangrove time-series products are constrained to 25 m resolution, hindering precise delineation of boundaries, small patches, and internal structures, thus compromising area estimates and ecological assessments. Key barriers are the paucity of historical high-resolution imagery and high-quality labeled samples. To this end, we developed the Sub-meter Mangrove Transfer Learning Mapping (SMTLM) framework, combining a pre-trained model from 2020 existing sub-meter products with transfer fine-tuning using few-shot samples from target historical years. We resampled time-series Landsat imagery to sub-meter resolution and selected proven spectral bands and vegetation indices. To improve temporal transferability, we designed and evaluated three strategies: direct transfer learning (DTL), partial fine-tuning (PFT), and layer-wise fine-tuning (LFT). SMTLM generated China's first national Sub-meter Time-series Mangrove Map (STMM) for 1990-2020, achieving overall accuracy of 93.4%-97.2% and F1-scores of 0.93-0.97. STMM indicates a 105% mangrove area increase from 2000 (14,316 ha) to 2020 (29,413 ha), nearly five times the 23.3% from existing product. When benchmarked against STMM, existing product exhibit commission and omission errors totaling 9,496–19,089 ha, and STMM detecting 2-4 times more mangrove patches. SMTLM provides a scalable solution for sub-meter mapping under sample-limited conditions, enhancing temporal transferability of high-resolution methods. STMM offers a long-term, high-resolution dataset for precise mangrove dynamics assessment, revealing far greater 21st-century recovery in China than previously recognized. Optimized 30 m Landsat imagery via SMTLM retrieves fine spatial details comparable to sub-meter results, highlighting sub-meter mapping's improvement over existing coarse-resolution products for reliable monitoring.