Dual Contrastive Network for Few-Shot Remote Sensing Image Scene Classification

Zhong Ji; Liyuan Hou; Xuan Wang; Gang Wang; Yanwei Pang

cs.CVMar 24, 2026

Dual Contrastive Network for Few-Shot Remote Sensing Image Scene Classification

Authors: Zhong Ji, Liyuan Hou, Xuan Wang, Gang Wang, Yanwei Pang

Organizations: School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China · CETC Key Laboratory of Aerospace Information Applications, Shijiazhuang, 050081, China

Abstract

Few-shot remote sensing image scene classification (FS-RSISC) aims at classifying remote sensing images with only a few labeled samples. The main challenges lie in small inter-class variances and large intra-class variances, which are the inherent property of remote sensing images. To address these challenges, we propose a transfer-based Dual Contrastive Network (DCN), which incorporates two auxiliary supervised contrastive learning branches during the training process. Specifically, one is a Context-guided Contrastive Learning (CCL) branch and the other is a Detail-guided Contrastive Learning (DCL) branch, which focus on inter-class discriminability and intra-class invariance, respectively. In the CCL branch, we first devise a Condenser Network to capture context features, and then leverage a supervised contrastive learning on top of the obtained context features to facilitate the model to learn more discriminative features. In the DCL branch, a Smelter Network is designed to highlight the significant local detail information. And then we construct a supervised contrastive learning based on the detail feature maps to fully exploit the spatial information in each map, enabling the model to concentrate on invariant detail features. Extensive experiments on four public benchmark remote sensing datasets demonstrate the competitive performance of our proposed DCN.

Figures & tables

Fig. 1: Illustrations of remote sensing image characteristics: small inter-class variances and large intra-class variances. See texts for details.

Fig. 2: Overview of the proposed DCN framework. It consists of a Context-guided Contrastive Learning (CCL) branch and a Detail-guided Contrastive Learning (DCL) branch to facilitate the model to learn more discriminative and detail information to generate context and detail features, both of which are used for training their corresponding classifiers.

Fig. 3: Illustration of Condenser Network. With the Squeeze and Expand operations among channels, the scene-relevant information is enhanced while the irrelevant information is ignored. After the global average pooling, the context features are generated.

Fig. 4: Illustration of Smelter Network. The helpful information in the channel and spatial dimensions is extracted respectively to enhance the significant local detail regions of the feature maps.

Type	Method	Backbone	WHU-RS19		UC Merced
Type	Method	Backbone	1-shot	5-shot	1-shot	5-shot
Optimization	LLSR [ 5 ]	Conv	57.10	70.65	39.47	57.40
Optimization	MAML [ 6 ]	ResNet-12	59.19 $\pm$ 0.92	72.34 $\pm$ 0.75	50.23 $\pm$ 0.98	62.02 $\pm$ 0.72
Metric	DLA-MatchNet [ 7 ]	Conv	68.27 $\pm$ 1.83	79.89 $\pm$ 0.33	53.76 $\pm$ 0.62	63.01 $\pm$ 0.51
	SCL-MLNet [ 8 ]	Conv	-	-	51.37 $\pm$ 0.79	68.09 $\pm$ 0.92
	MatchingNet [ 9 ]	ResNet-12	75.99 $\pm$ 0.63	82.96 $\pm$ 0.40	50.99 $\pm$ 0.69	57.14 $\pm$ 0.59

TABLE I: Comparative results(%) on the WHU-RS19 and UC Merced datasets with

\pm

\%

confidence intervals. The top two results are shown in bold and underline.

Type	Method	Backbone	NWPU-RESISC45		AID
Type	Method	Backbone	1-shot	5-shot	1-shot	5-shot
Optimization	LLSR [ 5 ]	Conv	51.43	72.90	45.18	61.76
Optimization	MAML [ 6 ]	ResNet-12	56.01 $\pm$ 0.87	72.94 $\pm$ 0.63	62.54 $\pm$ 0.79	73.12 $\pm$ 0.85
Metric	DLA-MatchNet [ 7 ]	Conv	68.80 $\pm$ 0.70	81.63 $\pm$ 0.46	57.21 $\pm$ 0.82	73.45 $\pm$ 0.61
	SCL-MLNet [ 8 ]	Conv	62.21 $\pm$ 1.12	80.86 $\pm$ 0.76	59.46 $\pm$ 0.96	76.31 $\pm$ 0.68
	MatchingNet [ 9 ]	ResNet-12	61.57 $\pm$ 0.49	76.02 $\pm$ 0.34	64.30 $\pm$ 0.46	74.49 $\pm$ 0.35

TABLE II: Comparative results(%) on the NWPU-RESISC45 and AID datasets with

\pm

\%

confidence intervals. The top two results are shown in bold and underline.

Method	WHU-RS19		UC Merced
Method	1-shot	5-shot	1-shot	5-shot
Backbone	67.65 $\pm$ 0.74	83.17 $\pm$ 0.36	51.87 $\pm$ 0.70	68.90 $\pm$ 0.55
CCL (w/o CL)	68.17 $\pm$ 0.72	84.08 $\pm$ 0.38	52.52 $\pm$ 0.72	70.02 $\pm$ 0.52
CCL (w/o CN)	67.13 $\pm$ 0.66	83.80 $\pm$ 0.37	54.37 $\pm$ 0.75	72.40 $\pm$ 0.48
CCL	69.03 $\pm$ 0.67	85.22 $\pm$ 0.40	54.93 $\pm$ 0.68	72.86 $\pm$ 0.50
DCL (w/o CL)	67.61 $\pm$ 0.69	83.11 $\pm$ 0.36	52.10 $\pm$ 0.65	69.57 $\pm$ 0.58

TABLE III: Ablation studies(%) on the WHU-RS19 and UC Merced datasets.

Method	NWPU-RESISC45		AID
Method	1-shot	5-shot	1-shot	5-shot
Backbone	71.26 $\pm$ 0.86	86.19 $\pm$ 0.47	69.22 $\pm$ 0.79	83.48 $\pm$ 0.45
CCL (w/o CL)	71.66 $\pm$ 0.85	86.31 $\pm$ 0.48	69.35 $\pm$ 0.84	83.07 $\pm$ 0.49
CCL (w/o CN)	72.01 $\pm$ 0.82	87.20 $\pm$ 0.44	69.18 $\pm$ 0.79	84.46 $\pm$ 0.45
CCL	72.55 $\pm$ 0.71	87.92 $\pm$ 0.41	70.91 $\pm$ 0.80	85.66 $\pm$ 0.44
DCL (w/o CL)	71.91 $\pm$ 0.75	86.53 $\pm$ 0.47	69.65 $\pm$ 0.82	84.10 $\pm$ 0.51

TABLE IV: Ablation studies(%) on the NWPU-RESISC45 and AID datasets.

Fig. 5: The inter-class variances and intra-class variances on the NWPU-RESISC45 dataset on 5-way 1-shot with baseline (ResNet), Condenser Network-integrated (ResNet+CN) and Smelter Network-integrated (ResNet+SN).

Fig. 6: Grad-CAM visualization results of inter-class images with different networks: baseline (ResNet) and Condenser Network-integrated (ResNet+CN).

Fig. 7: Grad-CAM visualization results of intra-class images with baseline (ResNet) and Smelter Network-integrated (ResNet+SN), where the right of the red dotted line is the visualization results of feature maps from several channels with large weights extracted by ResNet+SN.

Panel of Figure 12 — Fig. 8: Impact of weighting coefficient $\alpha$ on the NWPU-RESISC45 dataset.

Panel of Figure 13 — Fig. 9: Impact of weighting coefficient $\beta$ on the NWPU-RESISC45 dataset.

Panel of Figure 14 — Fig. 10: Impact of weighting coefficient $\gamma$ on the NWPU-RESISC45 dataset.

	WHU-RS19	UC Merced	NWPU-RESISC45	AID
Epoch	150	150	200	150
Time(s)	403.50	883.50	21010.00	4778.25

TABLE V: Training runtime on the WHU-RS19, UC Merced, NWPU-RESISC45 and AID datasets.

Remote Sensing Scene Classification

See all 11 papers

Locally Consistent Transductive Information Maximization for Few-Shot Remote Sensing Scene Classification
Jul 31, 2026Karim El Khoury, Benoît Gérin, Benoît Macq +1Remote Sensing Image Understanding Vision Foundation Model Adaptation
ICTEAM, UCLouvain, Louvain-la-Neuve, Belgium
Exploring the Potential of Contrastive Language-Image Pre-training for Multi-Source Remote Sensing Data
Sep 3, 2026Xiangyang Miao, Kelu Yao, Yekai Huang +7Remote Sensing Image Understanding Remote Sensing VLMs
School of Computer Science and Technology, Zhejiang University · Space-based Computing System Research Center, Zhejiang Lab
FROST: Training-Free Few-Shot Segmentation with Frozen Features and Nonparametric Statistics
Jun 30, 2026Junghwan ParkRemote Sensing Image Segmentation Few-Shot Semantic Segmentation
TelePIX

Jul 31, 2026J/K move · Enter open · S save

Locally Consistent Transductive Information Maximization for Few-Shot Remote Sensing Scene Classification

Karim El Khoury, Benoît Gérin, Benoît Macq, Christophe De Vleeschouwer

ICTEAM, UCLouvain, Louvain-la-Neuve, Belgium

Remote Sensing Image Understanding Vision Foundation Model Adaptation Remote Sensing Scene Classification Transductive Learning Remote Sensing Benchmarks Few-Shot Image Classification+1

Open paper AskPDF arXiv

Remote sensing scene classification is increasingly relying on foundation models pre-trained on large-scale Earth-observation data. Moreover, transductive inference, which exploits the collective statistical structure of the entire unlabeled query set, appears to naturally match remote sensing pipelines where large images are routinely split into patches and inferred as a batch. In this work, we introduce LC-TIM (Locally Consistent Transductive Information Maximization), which extends the state-of-the-art Transductive Information Maximization for Few-Shot CLIP (TIM++) objective with a local consistency regularizer that enforces prediction agreement between each query sample and its

κ

nearest feature-space neighbors. The regularizer enters as a single multiplicative factor in the closed-form

q

-update, adding negligible computational overhead. We further propose a multi-source extension that fuses the affinity graph from multiple remote sensing foundation model, further boosting classification accuracy. To assess these methods, we establish the first comprehensive, open-source benchmark for transductive few-shot RS scene classification, evaluating LP++, TransCLIP, TIM++, and LC-TIM across ten diverse datasets, two remote sensing vision-language models, and across various few-shot settings. Our experiments show that transductive methods consistently outperform zero-shot baselines, and that LC-TIM achieves state-of-the-art accuracy, with the largest gains in the low-shot regime where neighborhood cues are most informative. Code is publicly available at: https://github.com/elkhouryk/LC-TIM

Abstract

Remote Sensing Scene Classification

Explore similar work

Locally Consistent Transductive Information Maximization for Few-Shot Remote Sensing Scene Classification

Exploring the Potential of Contrastive Language-Image Pre-training for Multi-Source Remote Sensing Data

FROST: Training-Free Few-Shot Segmentation with Frozen Features and Nonparametric Statistics