Fine-scale forest classification with multi-temporal sentinel-1/2 imagery using a temporal convolutional neural network

Monitoring and mapping forest vegetation are crucial for conserving biodiversity and estimating biomass and carbon. However, spectral similarity between different vegetation types and the issue of mixed pixels in medium-resolution satellite imagery remain significant challenges for fine-scale forest classification. This study focused on the Tao River National Nature Reserve in southern Gansu Province, China, where over 6,000 vegetation samples were collected. A Temporal Convolutional Neural Network (TempCNN), which excels in handling temporal data and adapting to complex patterns, was constructed using TensorFlow with 133.8k parameters. The model was compared with CNN, random forest, XGBoost, and long short-term memory to validate its advantages. The SHAP (SHapley Additive exPlanations) interpreter was employed for feature importance analysis. Spectral, SAR, temporal, and geographic features were extracted from Sentinel-1/2 images to capture spectral differences and phenological characteristics of vegetation. Results show that TempCNN achieves the best performance (overall accuracy = 91.67%, Kappa = 0.895) and is particularly effective at handling complex data and mixed forests. Spring imagery and the short-wave infrared band play key roles in accurate vegetation classification. This study provides a new framework for fine-scale forest classification by combining multisource and multi-temporal data, highlighting the practical value of large-scale forest mapping.