基于长短时记忆网络的高保真遥感影像空谱联合分类方法

Spatial spectral joint classification of high fidelity remote sensing images based on long short memory network

摘要:

在遥感影像中,有些地物在光谱特征上的相似度较高,单纯依靠光谱信息难以实现准确区分。且高保真遥感影像包含大量的细节和丰富的信息,往往难以实现高效处理并挖掘信息之间的内在联系来捕捉地物的细微特征和上下文信息,导致在分类中,易受到噪声和各种干扰因素的影响,无法突出重要特征,使得不同类别区分效果不佳,造成分类精度较差。因此,提出基于长短时记忆网络的高保真遥感影像空谱联合分类方法。利用长短时记忆网络模型,分析高保真遥感影像特征,在此基础上,针对性处理目标影像数据,实现基于长短时记忆网络的高保真遥感影像压缩、感知与重构,以在有效减少数据量的同时,挖掘信息之间的内在联系来捕捉地物的细微特征和上下文信息。为抑制噪声和干扰信息,构建超像素小块,并根据小块内像素数据的降维运算标准,完善增强处理原则,突出重要特征,进而同时利用地物的空间与光谱信息定义空谱分类面,以有效区分不同类别,实现遥感影像的空谱联合分类,完成基于长短时记忆网络的高保真遥感影像空谱联合分类方法的设计。实验结果表明,光谱信息方面上述方法对目标地物遥感影像的分类精度更高；空间信息方面上述方法对遥感影像区域的划分更符合实际应用需求。

Abstract:

In remote sensing images, some land features have high similarity in spectral characteristics, and it is difficult to accurately distinguish them solely based on spectral information. Moreover, high fidelity remote sensing images contain a large amount of details and rich information, which often makes it difficult to achieve efficient processing and explore the inherent connections between information to capture subtle features and contextual information of ground objects. This leads to the susceptibility to noise and various interference factors in classification, making it difficult to highlight important features and resulting in poor classification accuracy for different categories. Therefore, a high fidelity remote sensing image spatial spectrum joint classification method based on long short-term memory network is proposed. By utilizing the Long Short Term Memory (LSTM) network model, the high fidelity remote sensing image features are analyzed. Based on this, targeted processing of target image data is carried out to achieve compression, perception, and reconstruction of high fidelity remote sensing images based on LSTM networks. This approach effectively reduces the amount of data while mining the inherent connections between information to capture subtle features and contextual information of ground objects. To suppress noise and interference information, superpixel blocks are constructed, and based on the dimensionality reduction operation standards of pixel data within the blocks, the enhancement processing principles are improved to highlight important features. Then, the spatial and spectral information of the land cover is used to define the spatial spectral classification surface, effectively distinguishing different categories and achieving spatial spectral joint classification of remote sensing images. The design of a high fidelity remote sensing image spatial spectral joint classification method based on long short-term memory networks is completed. The experimental results show that the above methods have higher classification accuracy for remote sensing images of target objects in terms of spectral information; In terms of spatial information, the above methods are more in line with practical application needs for the division of remote sensing image areas.