Remote sensing image classification is based on attention mechanism. How much do you know about algorithms?

The traditional pixel-based remote sensing image processing method is based on the rich spectral information of remote sensing images and the obvious spectral differences between objects. For high-resolution remote sensing images with only a few bands, traditional classification methods will lead to low classification accuracy and a large number of spatial data redundancy, and the classification results are often salt and pepper images, which is not conducive to spatial analysis. In order to solve this traditional problem, fuzzy classification technology came into being. Fuzzy classification is an image classification technology, which transforms the eigenvalues of any range into fuzzy values between 0 and 1, indicating the degree of belonging to the specified category.

By transforming eigenvalues into fuzzy values, fuzzy classification can standardize eigenvalues, even if they are combinations of eigenvalues with different ranges and dimensions. Fuzzy classification also provides a clear and adjustable feature description. For image classification, pixel-based information extraction is to classify each pixel according to the average radiation value within a certain pixel range on the surface. This classification principle makes a single pixel in high-resolution data or data with obvious texture characteristics of little value. In many cases (high-resolution or texture image data), the classification features of ground objects are expressed by texture features, not just spectral information.

In addition, all these background information is very important in image analysis. For example, urban green space and some wetlands are quite similar in spectral information. As long as the background of urban green space is defined as urban area in object-oriented image analysis, it is easy to distinguish green space from wetland, and this background information is hardly used in pixel-based classification. Object-oriented image analysis technology is the product of long-term development of spatial information technology, which has great potential in remote sensing image analysis.

So far, the object-oriented method is an ideal method, which can build a ground model that matches the real world. The most important part of object-oriented processing method is image segmentation. With the gradual refinement of earth observation tasks, high-resolution remote sensing satellite images have been more and more widely used. This brings challenges to the classification method of remote sensing images. The existing research shows that there are obvious limitations in the classification of high-resolution remote sensing images based on pixels. In recent years, Object-Based Image Analysis (OBIA) is considered as an important development trend of remote sensing and geographic information science, and it plays an increasingly prominent role in high-resolution remote sensing image processing.