Remote Sensing Semi-supervised Feature Extraction Framework And Lightweight Method Integrated With Distribution-aligned Sampling

In recent years, the semi-supervised element extraction task in remote sensing, which utilizes unlabeled data to assist training with a small amount of labeled data, has been widely explored. Most existing approaches adopt self-training or consistency regularization methods to enhance element extraction performance. However, there still exists a significant discrepancy in accuracy among different categories due to the imbalanced distribution of data classes. Therefore, a feature extraction Framework Integrated with Distribution-Aligned Sampling (FIDAS) framework is proposed in this paper. By leveraging historical data class distributions, the framework adjusts the training difficulty for different categories while guiding the model to learn the true data distribution. Specifically, it utilizes historical data distribution information to sample from each category, increasing the probability of difficult-category instances passing through thresholds and enabling the model to capture more features of difficult categories. Furthermore, a distribution alignment loss is designed to improve the alignment between the learned category distribution and the true data category distribution, enhancing model robustness. Additionally, to reduce the computational overhead introduced by the Transformer model, an image feature block adaptive aggregation network is proposed, which aggregates redundant input image features to accelerate model training. Experiments are conducted on the remote sensing element extraction dataset Potsdam. Under the setting of a 1/32 semi-supervised data ratio, a 4.64% improvement in mean Intersection over Union (mIoU) is achieved by the proposed approach compared to state-of-the-art methods. Moreover, while the essential element extraction accuracy is maintained, the training time is reduced by approximately 30%. The effectiveness and performance advantages of the proposed method in semi-supervised remote sensing element extraction tasks are demonstrated by these results. © 2024 Science Press. All rights reserved.