Enhanced mapping of mires in central Finland: integrating radar and optical data with machine learning

Under the dual pressures of global climate change and intensified human activities, peatlands, as one of the world’s most important terrestrial carbon reservoirs, are undergoing severe degradation and rapid area loss. Accurate and high-resolution peatland maps are therefore urgently needed to support peatland management, conservation, and global carbon monitoring. However, existing peatland mapping products generally suffer from limitations such as insufficient spatial resolution, incomplete spatial coverage, and long update cycles. To address these issues, this study developed a high-resolution peatland mapping framework by integrating Sentinel-1 synthetic aperture radar (SAR) data and Sentinel-2 multispectral instrument (MSI) data on the Google Earth Engine (GEE) cloud platform. A total of 15 optical features, 4 texture features, 15 radar features, and 2 phenological features were extracted from multi-source remote sensing imagery. Based on different combinations of these feature sets, six classification models were constructed and evaluated using the Random Forest (RF) algorithm. The study area covered two subregions in central Finland with a total area of 3 282 km², and a 10 m spatial resolution peatland map was generated. The results demonstrated that the integration of optical, radar, texture, and phenological features substantially improved peatland classification performance. The best-performing model achieved an overall accuracy of 94.7% and a Kappa coefficient of 89.8%. The inclusion of derived features effectively enhanced the separability between peatland and non-peatland classes and significantly improved the robustness and reliability of the classification results. Feature importance analysis revealed that the shortwave infrared band ratio (SWIR1/SWIR2) was the most effective discriminative variable, highlighting soil moisture conditions as a key indicator for distinguishing peatlands from surrounding land-cover types. In addition, this study proposed an innovative dual-polarization phenological index (VHND/VVND), which significantly enhanced the utilization efficiency of SAR data and improved the sensitivity of radar signals to seasonal hydrological variations. The results further indicated that the hydrological phenological period from May to June plays a decisive role in peatland identification in boreal regions. Compared with widely used global peatland products such as the Harmonized World Soil Database (HWSD) and PEATMAP, the proposed method overcomes limitations related to spatial resolution, temporal updating, and regional applicability. By integrating multi-source remote sensing data and phenological information within a cloud-computing environment, this study achieved both large-scale mapping capability and high classification accuracy simultaneously. The proposed framework provides a transferable and scalable technical approach for peatland mapping and offers important support for global peatland carbon stock assessment, climate change mitigation, and ecosystem conservation.