研究目的
To develop a high efficient deep feature extraction and classification method for spectral-spatial hyperspectral images that utilizes multiscale spatial features from guided filters to improve classification accuracy.
研究成果
The proposed MSCNN method with multiscale spatial features from guided filters significantly improves HSI classification accuracy and stability compared to state-of-the-art methods. It effectively addresses overfitting through regularization and dropout, and the use of multiscale features enhances spatial context utilization. Future work could explore automated parameter tuning and application to larger datasets.
研究不足
The method relies on limited labeled samples, which may affect generalization. The guided filter parameters (e.g., radius) are empirically set and might not be optimal for all datasets. The CNN model is simple and may not capture complex features as well as deeper networks, but deeper networks were found to be unsuitable due to overfitting with small datasets.
1：Experimental Design and Method Selection:
The methodology involves using a convolutional neural network (CNN) with dropout and L2 regularization to prevent overfitting, combined with a guided filter to extract multiscale spatial features from hyperspectral images (HSIs). The HSI spectral data is reshaped into 2D images for CNN processing.
2：Sample Selection and Data Sources:
Three benchmark datasets are used: Indian Pines, Pavia University, and Salinas, obtained from public sources. 10% of each dataset is used for training, and 90% for testing.
3：List of Experimental Equipment and Materials:
A PC with 32 GB memory, 8-core CPUs, and a GTX1060 GPU for acceleration. Software includes Python, Keras library, libSVM, and AE library.
4：Experimental Procedures and Operational Workflow:
Steps include data preprocessing (PCA for dimensionality reduction, reshaping spectra into 2D images), applying guided filter with different radii (r=3,5,7) to extract spatial features, fusing spectral and spatial features, training the CNN model with specified architecture, and evaluating performance using accuracy metrics.
5：Data Analysis Methods:
Performance is evaluated using overall accuracy (OA), average accuracy (AA), and kappa coefficient (KA). Statistical analysis includes comparing results with baseline methods like SVM and AE.
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
