FALSE: False Negative Samples Aware Contrastive Learning for Semantic Segmentation of High-Resolution Remote Sensing Image

The existing SSCL of RSI is built based on constructing positive and negative sample pairs. However, due to the richness ground objects complexity contextual semantics, same patches have coexistence imbalance samples, which causing pushing samples far away while away, vice versa. We call this confounding issue (SCI). To solve problem, we propose a False negAtive sampLes aware contraStive lEarning model (FALSE) for semantic segmentation high-resolution RSIs. Since pretraining unsupervised, lack definable criteria false (FNS) leads theoretical undecidability, designed two steps implement FNS approximation determination: coarse determination precise calibration FNS. achieve by self-determination (FNSD) strategy confidence (FNCC) loss function. Experimental results three datasets demonstrated that FALSE effectively improves accuracy downstream task compared with current models, represent different types models. mean Intersection-over-Union ISPRS Potsdam dataset improved 0.7\% average; CVPR DGLC 12.28\% Xiangtan 1.17\% average. This indicates has ability self-differentiate mitigates SCI in self-supervised contrastive learning. source code available at https://github.com/GeoX-Lab/FALSE.