Combined gated graph convolution neural networks with multi-modal geospatial data for forest type classification

Forest type classification is essential for the monitoring and management of forests, with significant implications for environmental protection and the mitigation of climate change. However, challenges such as multiscale variations, heterogeneous boundaries, mountainous terrain, and unbalanced data sets hinder progress. This study aims to improve forest type classification through three approaches: (1) Multimodal geospatial data fusion; (2) transfer learning using the ImageNet 22K dataset to improve accuracy and address class imbalances; (3) a novel Gated Graph Convolution Neural Network (GGCN). Experiments were conducted at two study sites with varying tree species, management strategies, and climates. The results indicated that very high-resolution aerial photographs outperform open-source Sentinel-1 and Sentinel-2 datasets. The fusion of the original remote sensing bands with the Enhanced Vegetation Index (EVI) feature demonstrates the best composition across all datasets. This approach, which combines the original Sentinel-1 and Sentinel-2 bands with the EVI, significantly improves the performance of open-source remote sensing data sets. It provides a cost-effective alternative to expensive high-resolution images, which is particularly beneficial for rural areas and global applications. Furthermore, utilizing ImageNet 22K transfer learning improved accuracy in addressing class imbalances. The GGCN effectively preserved multiscale and spatial features at both study sites. In general, this integrated approach shows promising potential for achieving high precision in large-scale forest type classification. • Forest type classification accuracy was refined through a novel deep learning model and multi-modal geospatial fusion datasets for two forests in Japan with different tree species composition, climate zones, and management strategies. • Multi-modal fusion breaks the limitations of single sensor and effectively improve classification accuracy. • Transfer learning significantly enhanced the classification accuracy and helped address the class imbalance problem.