Key Words: Deep Forest; Edge Detection; Multi-feature Fusion; Hyperspectral Image Classification.

Hyperspectral image (HSI) is widely used in many fields, since, it contains a lot of information. Relevant ideas in the field of computer vision are introduced into the task of feature extraction of hyperspectral images, and a deep learning framework is used to combine visual features with deep abstract features and needed for this use efficient feature learning methods to improve the classification effect. Multi-feature extraction and fusion strategy is an important tool to improve hyperspectral image classification, which can extract more information from HSI. There has been a lot of research for HSI classification, which are completely inability to use the spatial feature. However, the multi-feature extraction and feature fusion is not extensive, as a result, some HSI spatial features may not be fully extracted. Moreover, the fusion of features can easily lead to data redundancy. At the same time, the "Hughes" phenomenon and the band redundancy remain two important challenges in the coming decades. Relevant ideas in the field of computer vision are introduced into the task of feature extraction of hyperspectral images, and a deep learning framework is used to combine visual features with deep abstract features and use efficient feature learning methods to improve the classification effect. To overcome the lack of the mentioned classification the attention mechanism-based multi-feature fusion forest (Multi-Forest) method is proposed. The main work of this article includes:

(1)  The present multi-feature extraction method uses traditional methods to extract multiple shallow features to prove the use of HSI data by ensuring feature diversity. Such as Extended Morphological Profile (EMP), extract spatial information are used to extract spectral features. results have shown that the method can more effectively extract information about features. After reducing the original hyperspectral data, Extended Morphological Profiles is used to extract morphological features. The pixel structure reconstruction and the "erosion-dilation" operations are performed using structural elements of different sizes to achieve multiple "Open" and "Close" operation. After that, Strategies for extracting visual features are top-down techniques and consists of down-top strategies. The top-down technique for calculating visual salience by specific feature like previous knowledge and expected object is based on basic knowledge of human and down-top strategy is based on the automatic analysis of datasets by the Visual Silence model. VSD is then used to extract more important information in EMP images. For edge data loss due to VSD. Finally, Edge detection (ED) has been introduced feature fusion is performed by weighting, and then input to the classifier for classification.

(2) Deep learning methods have become the first choice for hyperspectral image (HSI) classification to date. Hyperspectral image classification based deep neural networks being applied and achieved great results. However, such approaches are still hampered by long training times. such as a large number of training samples, difficulty in adjusting the parameters, and easy over-fitting of the model, a framework of hyperspectral image classification based on attention mechanism based deep forest is proposed. First, by using a random forest, the feature importance of a spectral vector is calculated, which is used as an attention module to evaluate and improve the ability to express features. After that, the attention module is embedded in the deep forest, and the attention enhancement operation is performed on the spectral vector so that the local features of interest are strengthened during the classification process, and the purpose of adaptive feature refinement is achieved. Finally, the classification performance evaluation is performed by using the reduced-spectrum hyperspectral data and the fusion data based on visual features. The feature learning and image classification tasks are integrated into one to achieve the efficient classification of hyperspectral images.