研究目的
To tackle the problem of scarcity of labeled samples in scene classification for high-spatial resolution remote sensing images by proposing a novel semisupervised method that combines ResNet for feature extraction and ensemble learning to improve classification accuracy with limited labeled data.
研究成果
The proposed semisupervised method combining ResNet and ensemble learning achieves superior classification accuracy on multiple remote sensing image datasets, demonstrating effectiveness in leveraging unlabeled data to overcome label scarcity. Future work should focus on optimizing computational complexity and hardware improvements.
研究不足
The computational cost is higher due to training multiple base classifiers in ensemble learning; performance may degrade with very high proportions of labeled data due to noise from unlabeled samples; parameters T, r, and n need careful tuning to avoid overfitting or underperformance.
1：Experimental Design and Method Selection:
The method uses a pretrained ResNet50 to extract preliminary features from images, followed by an ensemble learning strategy with max-min sampling to create prototype sets from labeled and unlabeled data. Logistic regression is used for classification.
2：Sample Selection and Data Sources:
Datasets include AID (10,000 images, 30 classes), Sydney (1098 images, 8 classes), UC-Merced (2100 images, 21 classes), and WHU-RS (950 images, 19 classes), with labeled samples randomly selected in proportions (e.g., 1%, 2%, 5%, 8%).
3：List of Experimental Equipment and Materials:
Software tools include Keras with TensorFlow for ResNet implementation, LIBLINEAR and LIBSVM for classifiers, and sklearn for label propagation.
4：Experimental Procedures and Operational Workflow:
Images are fed into ResNet50 to get features; ensemble features are extracted using max-min sampling and logistic regression; classification is performed with labeled data.
5：Data Analysis Methods:
Performance is evaluated using average overall accuracy (ACC) from ten independent runs, with comparisons to state-of-the-art methods.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
