Improving semi-supervised remote sensing scene classification via Multilevel Feature Fusion and pseudo-labeling

Remote Sensing Image (RSI) scene classification plays a pivotal role in diverse applications such as land cover mapping, urban planning, and environmental monitoring. Traditional deep learning approaches, however, heavily rely on abundant labeled data, which is often costly and challenging to acquire. Semi-supervised learning emerges as a cost-effective alternative, yet existing methods frequently overlook the intricate characteristics of remote sensing data, such as multi-scale features and complex spatial patterns, thereby limiting their ability to effectively address these challenges. In this study, a novel Scene Semi-Supervised Method (SSSM) is introduced, marking a significant advancement in both network architecture and semi-supervised techniques. At the core of the SSSM framework lies the Multi-Level Feature Fusion Network (MFFN), meticulously designed to extract and integrate complex features from remote sensing data across diverse scales and locations. To optimize the utilization of pseudo-labels and minimize mislabeling, the Pseudo-Label Multi-Level Sampling strategy (PMLS) is proposed, a probabilistic approach that selectively identifies high-quality pseudo-labels to enhance training. Rigorous experiments conducted on three benchmark datasets reveal that the SSSM method significantly improves classification accuracy, achieving an increase of 3%–5% on a specific dataset compared to existing approaches. This accomplishment underscores the effectiveness of the MFFN design and the semi-supervised strategy in tackling the complexities of remote sensing scene classification. In summary, the MFFN-driven pseudo-label framework presented in this research pioneers a cutting-edge and promising new direction for semi-supervised remote sensing scene classification. • MFFN captures crucial features for scene classification with limited labeled data. • Introduced PMLS to select high-quality samples from pseudo-label data. • Attained 3%–5% average accuracy improvement over current methods in three datasets.