A comparative study of multi-label supervised contrastive losses for the content-based image retrieval of remote sensing images

Contrastive loss has been extensively studied for both supervised and unsupervised learning, and its success has led to its extension towards multi-label classification scenarios. Its application to multi-label content-based image retrieval however has been very limited. This study provides a systematic comparison in the context of content based remote sensing image retrieval where ever-growing data collections require efficient and effective management tools. Four multi-label supervised contrastive losses were investigated along with three benchmark datasets: BigEarthNet, UC Merced, and ML-AID. The weighted MulSupCon method achieved up to 90.11% mAP on ML-AID and 70.32% mAP on UC Merced, demonstrating its effectiveness in multi-label remote sensing retrieval tasks.