MACL: Multi-Label Adaptive Contrastive Learning Loss for Remote Sensing Image Retrieval

TL;DR

The paper tackles multi-label remote-sensing image retrieval under semantic overlap and label imbalance by extending supervised contrastive learning. It introduces MACL, which incorporates label-aware sampling, Pairwise Label Reweighting, and Dynamic Temperature Scaling to produce balanced, relationship-aware embedings. Across three benchmark datasets, MACL and its weighted variant consistently outperform existing contrastive baselines in both cosine-based and Jaccard-based retrieval metrics, with ablations confirming the contribution of each adaptive component. The results suggest MACL offers robust, scalable improvements for large-scale remote-sensing archives and motivates future work on higher-order label correlations and semi-supervised extensions.

Abstract

Semantic overlap among land-cover categories, highly imbalanced label distributions, and complex inter-class co-occurrence patterns constitute significant challenges for multi-label remote-sensing image retrieval. In this article, Multi-Label Adaptive Contrastive Learning (MACL) is introduced as an extension of contrastive learning to address them. It integrates label-aware sampling, frequency-sensitive weighting, and dynamic-temperature scaling to achieve balanced representation learning across both common and rare categories. Extensive experiments on three benchmark datasets (DLRSD, ML-AID, and WHDLD), show that MACL consistently outperforms contrastive-loss based baselines, effectively mitigating semantic imbalance and delivering more reliable retrieval performance in large-scale remote-sensing archives. Code, pretrained models, and evaluation scripts will be released at https://github.com/amna/MACL upon acceptance.

Figures (11)

• Figure 1: Illustration of the MulSupCon method. Each row represents a sample’s one-hot label vector, with the first row being the anchor sample (top row) and the remaining rows corresponding to other samples in the batch. Filled circles indicate active label dimensions. Samples enclosed with dotted outlines belong to the negative set $\mathcal{N}$, while the remaining samples form the positive set $\mathcal{P}$. For an anchor with labels $\{1,3\}$, the corresponding positive sets are $P_1 = \{1,3,6\}$ and $P_3 = \{1,2,3,4,6\}$, which form the basis for label-wise contrastive learning in MulSupCon.
• Figure 2: Training and retrieval pipeline employing the proposed MACL loss. During training, images are encoded by a shared backbone and projection head to produce embeddings optimized using the MACL loss. During retrieval, query and database images are embedded into a common feature space and ranked based on similarity.
• Figure 3: Visualization of the MACL mechanism. For each anchor label, MACL computes a label-specific gradient component using pairwise weights (green/yellow/red nodes) and adaptive temperatures (color bar). The effective influence of each positive is determined by $\gamma_{ip} = w_{ip}/T_{ip}$, and the final gradient is obtained by combining the label-specific components.
• Figure 4: Examples of data augmentations applied to the DLRSD dataset.
• Figure 5: Examples of data augmentations applied to the ML-AID dataset.