Polarimetric SAR surface classification

The traditional polarimetric image classification method usually consists of two parts: feature extraction and classifier design , The quality of polarization feature extraction plays a decisive role in the classification effect . Usually, a lot of manual design work is required in the feature extraction stage , This part of the work is not only time-consuming and laborious , And often requires professionals to SAR Have a deep understanding of the system , To design the right features . In such shallow networks , The generalization ability of classifier is not very strong . Designing classifiers also requires other skills . In polarization SAR Surface classification , The accuracy of classification is not only related to the information of polarization characteristics , It is also closely related to spatial features . therefore , The unique characteristics of deep convolution neural network classifier have important application potential in surface classification . Polarization characteristic information can usually be extracted from backscattered electromagnetic waves , Spatial information is usually not only related to the target itself , It is also related to the neighborhood information of the target . However , Deep convolution neural network can automatically carry out feature extraction and feature classification , It is a very good choice for polarimetric image classification .
Each hidden layer node of convolutional neural network can be regarded as a feature detector , When a feature that it detects appears in its input , This node has a large response value . All nodes on the same characteristic graph are restricted to share the same connection weight , Therefore, each feature map detects the same feature at different positions of the image .
The weight of convolution layer is also called convolution kernel , Visualizing these weights can help us better understand the working principle of convolution network , To understand the significance of feature detection .
notes ： The article is excerpted from 《 Synthetic aperture radar intelligent interpretation 》 Xu Feng et al