Texture based prototypical network for few-shot semantic segmentation of forest cover: Generalizing for different geographical regions

Forest plays a vital role in reducing greenhouse gas emissions and mitigating climate change, besides maintaining the world's biodiversity. The existing satellite-based forest monitoring system utilizes supervised learning approaches limited to a particular region and depends on manually annotated data to identify forest. This work envisages forest identification as a few-shot semantic segmentation task to achieve generalization across different geographical regions. The proposed few-shot segmentation approach incorporates a texture attention module in the prototypical network to highlight the texture features of the forest. Indeed, the forest exhibits a characteristic texture different from other classes, such as road, water, etc. In this work, the proposed approach is trained for identifying tropical forests of South Asia and adapted to determine the temperate forest of Central Europe with the help of a few (one image for 1-shot) manually annotated support images of the temperate forest. An IoU of 0.62 for forest class (1-way 1-shot) was obtained using the proposed method, which is significantly higher (0.46 for PANet) than the existing few-shot semantic segmentation approach. Besides, the experimental results demonstrate that the inclusion of the texture attention module in the existing prototypical few-shot segmentation methods (PFENet and ASGNet) results in a more accurate forest identification. These results indicate that the proposed approach can generalize across geographical regions for forest identification, creating an opportunity to develop a global forest cover identification tool.