Objective This paper aims to combine multi-source remote sensing data for feature extraction to determine the most effective classification strategy. Additionally, we investigated the significance of time series features in identifying forest types, offering a technical approach for remote sensing-based forest type identification.
Method This study combined Sentinel-2 spectral features, time series features, Sentinel-1 radar backscatter features, and SRTM DEM terrain features to extract various feature variables using Google Earth Engine. Multiple feature combinations were constructed and classified using the random forest classifier. Subsequently, mapping output and accuracy evaluations were performed on the resulting classifications.
Result (1) The scheme that incorporates Sentinel-2 time series features, Sentinel-1 radar backscatter features, and SRTM DEM terrain features exhibited the highest classification accuracy, achieving an overall accuracy of 84.62% and a Kappa coefficient of 0.82. (2) Among the five constructed feature combination schemes, the multi-feature combination scheme demonstrated superior classification performance compared with individual feature. (3) Terrain features, radar backscatter features, and time series features significantly influenced the classification results. The inclusion of time series features notably enhanced the accuracy of forest type identification. Among the spectral features, the shortwave infrared bands B11 and B12 were the most critical, while April and October were identified as the most important time nodes within the time series features.
Conclusion The multi-feature classification scheme, which combines data from various remote sensing sources, is proved to be effective in accurately identifying forest types in the study area. SRTM DEM terrain features, Sentinel-1 radar backscatter features, and Sentinel-2 time series features serve as valuable complementary indicators to spectral features, enhancing classification accuracy. Time series features, in particular, play a significant role in improving the accuracy of forest type identification.