Mapping tree species in a subtropical forest landscape: a deep learning approach using a composite time series index for active–passive remote sensing

Subtropical evergreen forests are highly complex. Similar spectral features of forest canopies and limited temporal densities of optical imagery pose significant challenges for tree species mapping. Focusing on the hinterland of China’s Ta-pieh Mountains, we used a limited annual Sentinel-1/2 time series dataset to propose an optical–SAR combined vegetation index (OSCVI) and develop a spectral–temporal frequency mixer enhanced Swin Transformer (STFM–SwinT). The results showed the following: (1) The time series OSCVI (TOSCVI) exhibited low redundancy, and its mapping accuracy (Kappa) reached 80.01% (0.75), improving by 3% (0.039)–5.38% (0.065) over purely optical features. (2) When texture features (TXT) were added to TOSCVI (TOSCVI+TXT), the highest mapping accuracy (Kappa) reached 83.3% (0.792), which was a 3.29% (0.042) improvement over the previous best result. (3) When the constructed active–passive features were used, the STFM–SwinT model achieved a mapping accuracy greater than that of six other time series deep learning models and two machine learning models, with improvements in accuracy (Kappa) of 3.58% (0.045)–7.03% (0.081) and 9.66% (0.103)–15.62% (0.176), respectively. This approach effectively addresses common issues such as “same spectrum, different objects” and the insufficient temporal density of optical images in subtropical tree species.