Application of the Vision Transformer and Mask R-CNN Joint Algorithm to Assist Forest Decisions

With the continuous iterative development of remote sensing technology and deep learning algorithms, it provides strong support for the rational development, management, and protection of forest resources. This article proposes a Vision Transformer and Mask R-CNN joint algorithm architecture that applies NEON high-resolution remote sensing image data to achieve forest type classification and recognition, as well as single tree crown area instance segmentation. Research has found that the ViT algorithm performs well in large-scale data, but it relies on reasonable parameter settings, and the ViT model pre-trained through a large dataset outperforms the vast majority of other models at present. Mask R-CNN relies on the accuracy of data annotation, and models trained with a small amount of data can also achieve good training results.