Robust Land Cover Classification With Local-Global Information Decoupling to Address Remote Sensing Anomalous Data.

Remote sensing images play a critical role in urban planning, land resources and environmental monitoring. Land cover classification is one of the straightforward applications of remote sensing. However, the anomalous remote sensing data challenges the reliability of land cover classification results. Deep learning has been widely used in remote sensing image analysis, but it remains sensitive to anomalous data. To address this issue, we re-evaluate a land cover classification map in high-noise scenarios with anomalous data and propose a novel network architecture to solve the problem. A new network architecture is proposed to solve this problem. Our proposed network architecture focuses on decoupling the extraction of global information and local information. Through three global-local feature fusion modules, we output features emphasizing global information, features emphasizing local information, and consistency evaluation scores, respectively. A specially designed decoder integrates these three features. Our method performs better compared to mainstream models on the public datasets WHDLD with obvious anomaly data, with a MIoU of 63.58% and a Mpa of 74.32%. Compared to the suboptimal method, our method improves MIoU by 1.29% and Mpa by 3.05%.