Hyperspectral image (HSI) classification often faces the problem of multi-class imbalance, which is considered to be one of the major challenges in the field of remote sensing. In recent years, deep learning (DL) has been successfully applied to HSI classification, convolutional neural network (CNN) is one of the most representative of them. However, it is difficult to effectively improve the accuracy of minority classes under the problem of multi-class imbalance. In addition, ensemble learning has been successfully applied to solve multi-class imbalance, such as random forest (RF). This paper proposes a novel ERFS (enhanced random feature subspace)-based ensemble CNN algorithm for the multi-class imbalanced problem. The main idea is to perform random oversampling (ROS) of training samples and multiple data enhancements based on random feature subspace (RFS), and then construct an ensemble learning model combining random feature selection and CNN to HSI classification. Experimental results on three public hyperspectral datasets show that the performance of the proposed method is better than traditional CNN, RF, and deep learning ensemble methods. Index Terms-Hyperspectral image (HSI) classification, multiclass imbalance, enhanced random feature subspace (ERFS), ensemble learning, convolutional neural network (CNN).
The fusion of multi-spectral and synthetic aperture radar (SAR) images could retain the advantages of each data, hence benefiting accurate land cover classification. However, some current image fusion methods face the challenge of producing unexpected noise. To overcome the aforementioned problem, this paper proposes a novel fusion method based on weighted median filter and Gram–Schmidt transform. In the proposed method, Sentinel-2A images and GF-3 images are respectively subjected to different preprocessing processes. Since weighted median filter does not strongly blur edges while filtering, it is applied to Sentinel-2A images for reducing noise. The processed Sentinel images are then transformed by Gram–Schmidt with GF-3 images. Two popular methods, principal component analysis method and traditional Gram–Schmidt transform, are used as the comparison methods in the experiment. In addition, random forest, a powerful ensemble model, is adopted as the land cover classifier due to its fast training speed and excellent classification performance. The overall accuracy, Kappa coefficient and classification map of the random forest are used as the evaluation criteria of the fusion method. Experiments conducted on five datasets demonstrate the superiority of the proposed method in both objective metrics and visual impressions. The experimental results indicate that the proposed method can improve the overall accuracy by up to 5% compared to using the original Sentinel-2A and has the potential to improve the satellite-based land cover classification accuracy.
Convolutional neural networks (CNNs) can automatically learn features from the hyperspectral image (HSI) data, avoiding the difficulty of manually extracting features. However, the number of training samples for the classification of HSIs is always limited, making it difficult for CNN to obtain effective features and resulting in low classification accuracy. To solve this problem, a pixel cluster CNN and spectral-spatial fusion (PC-CNN-SSF) algorithm for hyperspectral image classification (HSIC) with small-size training samples is proposed in this paper. Firstly, spatial information is extracted by the Gray Level Co-occurrence Matrix (GLCM). Then spatial information and spectral information are fused by means of bands superposition, forming spectral-spatial features. To expand the number of training samples, the pixels after spectral-spatial fusion are combined into pixel clusters according to a certain rule. Finally, a CNN framework is utilized to extract effective features from the pixel clusters. Experiments based on three standard HSIs demonstrate that the proposed algorithm can get better performance than the conventional CNN and also outperforms other studied algorithms in the case of small training set.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.