Temporal variability in remote sensing accuracy for wetland mapping: a case study from Sentinel-1 and Sentinel-2 in Northeast China

Wetlands are vital ecosystems that support regional ecological balance and require efficient, accurate, and cost-effective monitoring approaches. This study enhances wetland mapping in the Sanjiang Plain using multi-source (Sentinel-1 SAR and Sentinel-2 optical) and multi-temporal (2018–2021) satellite data on the Google Earth Engine platform. A Random Forest classifier was applied with training samples from high-resolution imagery and field surveys. Input features included spectral bands, vegetation indices, and radar backscatter coefficients. Preprocessing involved cloud masking for Sentinel-2 and speckle filtering for Sentinel-1. Classification accuracy was evaluated using independent validation samples, overall accuracy, and Kappa coefficient. Results indicate October images provide the highest single-year mapping accuracy, with Sentinel-1 and Sentinel-2 data from October 2021 achieving 85.4% accuracy and a Kappa of 0.669. Multi-temporal data improved accuracy to 88.9% (Kappa = 0.749). Sentinel-1 showed greater annual variation in wetland distribution compared to the stable Sentinel-2. Precipitation impacted accuracy, reducing Sentinel-2 performance while variably affecting Sentinel-1. Combining radar and optical data with multi-temporal analysis enhances wetland monitoring, offering guidance for data acquisition in large-scale conservation. Challenges include misclassifications like water-shadow confusion in optical imagery and backscatter interference from vegetation in radar data, requiring further research.