Visual recognition algorithm for weakly labeled multi-source data fusion of remote sensing images

To address the limitations of existing fine-grained visual recognition methods that rely heavily on strong annotations and exhibit poor cross-domain generalization, this study proposes a weakly annotated multi-source data fusion algorithm for remote sensing image recognition. The algorithm achieves three core objectives: efficient utilization of weakly annotated data, precise extraction of fine-grained features, and enhanced cross-domain generalization. First, we develop a self-supervised learning-based weak annotation processing framework with attention mechanisms. Through self-supervised tasks like rotation prediction and image restoration, we extract effective features from weakly annotated data. The attention mechanism is then employed to filter high-confidence samples, significantly improving data utilization efficiency. Second, we design a multimodal cross-attention fusion model that maps local image features and textual semantic features into a shared embedding space. By integrating cross-attention and gating mechanisms, we overcome modal gaps and enhance fine-grained feature recognition. Finally, we introduce cross-domain contrastive loss, combining contrastive learning with domain adaptation techniques to identify domain-invariant features across source and target domains, thereby improving cross-domain generalization performance. Experimental results demonstrate that the algorithm achieves a 94% classification accuracy on traditional datasets, outperforming existing baseline models while maintaining stable performance in real-world scenarios.