1. Introduction
Remote sensing image classification is crucial for environmental monitoring and urban planning, yet it faces challenges like data scarcity, domain shifts, and limited generalization of learned models. Traditional methods often struggle to fully exploit the rich semantic information available from multimodal sources, such as accompanying textual descriptions or labels. This work proposes to address these limitations by developing a network that can effectively fuse frequency-aware visual features with linguistic context. Models used in this article include the proposed FA-VLMGN, baseline models such as ResNet-50, Vision Transformer (ViT), and multimodal approaches like CLIP and its adaptations for remote sensing.
2. Related Work
Recent advancements in vision-language pre-training, exemplified by models like CLIP and ALIGN, have shown remarkable capabilities in zero-shot and few-shot classification across diverse domains. In remote sensing, efforts have been made to adapt these models, often focusing on aligning visual and textual embeddings. Concurrently, frequency domain analysis has proven effective in computer vision for noise robustness and feature enhancement, but its integration with multimodal vision-language models for generalization in remote sensing remains underexplored. This section reviews these separate lines of research and highlights the gap our proposed method aims to bridge.
3. Methodology
The proposed Frequency-Aware Vision-Language Multimodality Generalization Network (FA-VLMGN) begins by processing remote sensing images through a vision encoder that incorporates a frequency attention mechanism to extract robust visual features. Simultaneously, a language encoder processes textual labels to generate semantic embeddings. A novel cross-modal fusion module then aligns and integrates these frequency-aware visual and linguistic representations. Finally, a generalization head employs an adaptive weighting scheme to ensure the model's robustness and transferability across different remote sensing datasets and scenarios, minimizing domain-specific biases.
4. Experimental Results
Experiments conducted on several remote sensing benchmark datasets, including NWPU-RESISC45 and AID, demonstrate the superior performance of FA-VLMGN. The model achieved significantly higher classification accuracy and improved robustness against domain shifts compared to state-of-the-art methods. Ablation studies confirmed the critical role of both the frequency-aware module and the adaptive generalization mechanism in achieving these results. The following table illustrates a comparative performance on the NWPU-RESISC45 dataset, showcasing FA-VLMGN's advantage in overall accuracy over various baselines. The table presents the classification accuracy (in %) of different models on the NWPU-RESISC45 dataset, clearly indicating that the proposed FA-VLMGN significantly outperforms both unimodal and other multimodal baseline approaches. This improvement highlights the effectiveness of incorporating frequency-aware features and a robust generalization strategy for remote sensing image classification tasks, suggesting a more comprehensive understanding of complex aerial scenes.
5. Discussion
The enhanced performance of FA-VLMGN can be attributed to its ability to leverage frequency domain information, which provides more invariant and discriminative visual features, and its effective integration with rich linguistic semantics. The robust generalization capabilities observed suggest that the model can better adapt to unseen remote sensing scenarios, reducing the need for extensive retraining. This work not only advances the state-of-the-art in remote sensing image classification but also opens new avenues for developing more interpretable and robust multimodal AI systems in this domain. Future work could explore dynamic frequency selection and meta-learning strategies for even broader generalization.