Is there any good method for supervised classification of remote sensing images?

According to the samples provided by the known training area, the discriminant function is established by calculating and selecting characteristic parameters, and the images to be classified are classified.

Supervised classification, also known as training ground method, is a technique based on establishing statistical identification function and a typical sample training method. That is, according to the samples provided by the known training area, feature parameters are selected, found out as decision rules, and a discriminant function is established to classify the images to be classified, which is a pattern recognition method. The training area is required to be typical and representative. If the criterion meets the requirements of classification accuracy, the criterion is established; On the contrary, the decision rules of classification need to be re-established until the classification accuracy requirements are met. Commonly used algorithms include discriminant analysis, maximum likelihood analysis, feature analysis, sequence analysis and pattern recognition.

Process:

1. Select the training area (representative, complete and multi-sample area)

2. Extract statistical information (multivariate statistical analysis, effective evaluation of training samples, sample purification).

3. Choose the appropriate supervised classification algorithm (parallel algorithm, minimum distance method, maximum likelihood method (the most widely used so far), spectral angle classification).

4, computer automatic classification

5. Evaluation of classification accuracy (non-positional accuracy, positional accuracy-confusion matrix)

Advantages:

1, we can make full use of the prior knowledge of the classification area to determine the classification category in advance;

2. The selection of training samples can be controlled, and the training samples can be repeatedly tested to improve the classification accuracy and avoid serious errors in classification.

3. The re-classification of spectral clustering in unsupervised classification is avoided.

Disadvantages:

1, with strong subjective factors;

2. The selection and evaluation of training samples need more manpower time;

3. Only the categories defined in the training samples can be identified, thus affecting the classification results.