Data Fusion and Models Integration for Enhanced Semantic Segmentation in Remote Sensing

Remote sensing semantic segmentation is a key research area in the remote sensing domain. Despite advancements, there is still no unified standard dataset such as ImageNet for remote sensing segmentation. This fragmentation leads to scattered efforts, with different teams using various datasets, making horizontal comparisons and result sharing difficult. To address this issue, we propose a novel data fusion method for remote sensing semantic segmentation, which integrates remote sensing data from different sources and segmentation results, providing support for the creation of a unified, large-scale semantic segmentation dataset suitable for deep learning. This method promotes standardized research in remote sensing segmentation and provides a reference benchmark for cross-comparison of different research outcomes. To demonstrate the feasibility of the proposed method, we have constructed a fused dataset named GGDS, which integrates Google Maps satellite imagery with GlobeLand30 semantic labels. In addition, we designed and implemented model integration techniques for the fused dataset, improved the segmentation_models_pytorch library, and developed the EffiTUnet model based on it, achieving outstanding experimental results. The methods and tools presented in this article offer substantial support and convenience for related research. Our results not only account for the diversity and complexity of different remote sensing data but also enhance model generalization in various application scenarios through ensemble learning strategies, providing strong support for training more powerful and efficient semantic segmentation models.