Modeling study demonstrates enhanced tree species mapping accuracy using active-passive satellite time series, indicating improved monitoring of complex subtropical forest canopies.
Key Points
To overcome spectral similarities and optical data gaps when classifying tree species in complex subtropical evergreen forests.
Generated an optical–SAR combined vegetation index (OSCVI) and texture features (TXT) using annual Sentinel-1 and Sentinel-2 satellite time series.
Developed a spectral–temporal frequency mixer enhanced Swin Transformer (STFM–SwinT) and evaluated it against six deep learning and two machine learning models.
The time series OSCVI achieved a mapping accuracy of 80.01% (Kappa 0.75), exceeding purely optical features by 3.00% to 5.38% (Kappa improvements of 0.039 to 0.065).
Adding texture features (TOSCVI+TXT) achieved a peak classification accuracy of 83.30% (Kappa 0.792), an improvement of 3.29% (Kappa 0.042).
The STFM–SwinT model outperformed six time-series deep learning models by 3.58% to 7.03% (Kappa gains of 0.045 to 0.081) and two machine learning baselines by 9.66% to 15.62% (Kappa gains of 0.103 to 0.176).