PLANET: Multi-Class Patch Layer Adaptive Network for Satellite Image Segmentation

Region of interest segmentation is an important step in satellite image analysis. U-Net is a commonly adopted segmentation model, which shows good performance in many applications. However, applying U-Net to satellite images requires high memory usage, and is constrained to binary classification. To address these issues, we propose Patch Layer Adaptive Network (PLANET), which introduces dynamic layer-design and multi-class capabilities. We compared PLANET with a number of models using the MBRSC satellite dataset, and performed both qualitative and quantitative analysis. Experimental results demonstrated that PLANET outperformed other methods, achieving a mIoU score of 89.4% compared to U-Net's 79.4%.