Fusion of Multiple Models with Multi-Modal Datasets: Land Cover Mapping in the Yangtze River Economic Belt

Jiepan Li, He Huang, Yu Xia

Abstract: Land cover is essential for understanding Earth's environment, playing a vital role in disaster monitoring and infrastructure management. The Yangtze River Economic Belt, a pivotal region for economic growth in China, requires comprehensive monitoring and analysis of land cover changes to support effective policy-making and environmental protection. The 2024 IEEE Whisper contest, MMSeg-YREB, aims to integrate multi-modal data, specifically Sentinel-1 and Sentinel-2, to conduct a detailed analysis of land cover types and their changes within the Yangtze River Economic Belt. To overcome this challenge, the MMSeg-YREB dataset was integrated with Google Earth Engine (GEE) tools and publicly available global land cover products for data preparation. With this enhanced dataset, multiple models were developed, including convolutional neural networks (CNN) and Transformer networks, which enables the learning of features from both local and global perspectives, significantly enhancing the accuracy of segmentation. To fuse these results, a variety of post-processing techniques were proposed to refine the outcomes of the different models, resulting in a highly accurate final land cover maps. This innovative approach allowed our team to achieve a commendable first-place in the competition, underscoring our contributions and effectiveness in advancing methodologies for land cover mapping.