Remote sensing image Processing Learning note Two--classification processing of remote sensing images

1. Basic Concepts

Class of figure: a set of figures with the same characteristics is called a class.

A kind of figure has the same sign, and different kinds of objects have different spectral characteristics (the ability to reflect and emit electromagnetic energy)

Classification: According to the intrinsic similarity of each kind of sample, the process of dividing the feature space into several sets is adopted by some judgment criterion.

2. Basic ideas

The theoretical basis of distinguishing different figures: different figure types have different spectral information and spatial information.

3. Multi-Spectral Image representation method:

Assuming that a multispectral image includes n bands, the luminance values for pixels in either position (I, j) can be expressed in vector x= (x1, x2, ..., xn), where each component Xi represents the luminance value of the pixel in the first band.

The multispectral image can then be represented by a series of points in the N-dimensional feature space.

However, due to the various state of the figures and various interference factors in the imaging, the spectral response characteristics of each type of ground object obtained by the sensor are not identical.

Therefore, the same class of object samples, in the spectral space, the performance of the following forms:

A probability distribution or aggregation around a point.

4. Basic Process

Defining classification Categories--feature selection and extraction--training data--statistical characteristics evaluation--analysis of classification results

Spectral characteristics are widely used in the classification of remote sensing images.

5. Classification

{Unsupervised classification

{Cell-based classification

{Supervised classification

Image classification

{Object-oriented classification

Unsupervised classification: There is no prior category information, not known in advance into several categories, only according to the size of the similarity between the elements of the method of classification and merging;

Supervised classification: It is necessary to select some areas of training that can be clearly identified;

Sampling of training data in the region;

The statistical distribution characteristics of the object categories are obtained by using the sampled data.

The other pixels are classified according to the obtained characteristics.

The following is a simple way of understanding the K-mean algorithm for unsupervised classification:

Steps:

(1) The center position of a given n initial class;

(2) Calculate the distance from each pixel to the center of all classes and re-partition the pixel into the nearest class;

(3) updating various centres;

(4) Repeat (2) and (3) to know that the central position has not changed significantly.

Characteristics: Given the number of categories in advance (self-assessment), local optimal, dynamic clustering.

6. Object-oriented remote sensing image classification

(1) Firstly, the defects of remote sensing image classification based on cell are mentioned:

1) It is difficult to overcome the limitation of spectral information such as "same-spectrum foreign body" and "same-matter heterogeneous spectrum";

2) The value of the image's previous cell contains the signal from the neighboring cell corresponding to the surface.

(2) Process:

(3) Image segmentation:

Function: The image is divided into several "meaningful" disjoint regions, so that these characteristics are consistent or similar in a certain region, showing a distinct difference in different regions.

Principle: The main use of spectral features and shape features to adjust the object boundary.

The formula is not good hit here will not fight ...

The Split function is as follows:

There are two criteria, one is the spectral criterion and the other is the shape criterion. Is the weight of the spectrum relative to the shape, which is defined by the user itself.

The spectral criterion can be calculated by formula.

The shape criterion is calculated by the smoothness (smoothness) and the tightness (compactness) .

Smoothness is the ratio of the object's circumference l to the minimum bounding rectangle perimeter B: smooth=l/b.

The tightness is the ratio of the object's circumference L to the square root of the object's size (the number of pixels of the object): cpt=l/

The formula for the shape criterion is:

Where (between 0 and 1) is the user-defined weight of the tightness criterion. The criterion of tightness and the criterion of smoothness are calculated on the basis of the formula.

(4) Multi-scale segmentation

Definition: Start with a bottom-up region merging technique from a pixel object.

Preparation: User-specified criteria for spectral and shape parameters and neighborhood function logic.

Thought: By defining these criteria, the individual cells are expanded into homogeneous regions (the work performed by the image segmentation above).

This is a gathering process. Adjacent object pairs in each step are merged as long as the minimum growth criteria for the defined heterogeneity are met, and if the merge exceeds the threshold range defined by the scale parameter, the merge process stops. (Good ...) Good abstraction ... Embarrassing).

7. Ways to improve classification accuracy

Reasons for low classification accuracy:

(1) Low image resolution

(2) Classification features single

(3) Single classification method

For the reason that the classification method is single, the classification precision is low, and the different methods are suitable for different features. The workaround for this problem is:

(1) Integration of multiple classifiers: such as supervised and unsupervised integration;

(2) Adopt a new computer classification method: such as fuzzy classification method, artificial neural network Method (*), based on knowledge classification method.

For the method of fuzzy classification, the main idea is not to consider "sample x belongs to Class B", but to consider "how much the sample x belongs to Class B".

Remote sensing image Processing Learning note Two--classification processing of remote sensing images