SBSS: Stacking-Based Semantic Segmentation Framework for Very High-Resolution Remote Sensing Image

Semantic segmentation of Very High Resolution (VHR) remote sensing images is a fundamental task for many applications. However, large variations in the scales objects those VHR pose challenge performing accurate semantic segmentation. Existing networks are able to analyse an input image at up four resizing scales, but this may be insufficient given diversity object scales. Therefore, Multi Scale (MS) test-time data augmentation often used practice obtain more results, which makes equal use results obtained different it was found study that classes had their preferred scale Based on behaviour, Stacking-Based Segmentation (SBSS) framework proposed improve by learning contains learnable Error Correction Module (ECM) result fusion and Scheme (ECS) computational complexity control. Two ECS, i.e., ECS-MS ECS-SS, investigated study. The Floating-point operations (Flops) required ECS-SS similar commonly MS test Single-Scale (SS) test, respectively. Extensive experiments datasets (i.e., Cityscapes, UAVid, LoveDA Potsdam) show SBSS effective flexible framework. It achieved higher accuracy than when using ECS-MS, as SS with quarter memory footprint ECS-SS.