Accuracy Improvement of the Land Cover Classification by using Truncated Normal Distribution
4.1 Maximum likelihood classification using truncated normal distribution
The range of (µ??s?µ??s, µ: mean value, s: standard deviation) in normal distribution contains about 95%. Normal distribution was truncated in this range. Truncated normal distribution was created.
4.2 Maximum likelihood classification by the screening of the training data
There are two following types for the screening of the training data
? The pixel was removed when pixel value exist in the tail of training data distribution at least one band.
? The pixel is removed when pixel value exist in the tail of training data distribution at all bands
In this study, screening object is made to be the data in which the data does not come in within ( µ-2s,µ+2sµ : Mean value s : Standard deviation ) in each class. Maximum likelihood classification was implemented by using screening ? and screening ?.
4.3 Maximum likelihood classification using the both methods
Classification accuracy using both methods of the maximum likelihood classification of the training data using screening and truncated normal distribution proposed in this paper.
5. Result
Standard classification accuracy must be prepared to confirm the improvement of classification accuracy. Multisensor fusion image of the Landsat/TM 3 bands and SPOT panchromatic image was used. The maximum likelihood classification result for this image is used as the standard classification accuracy. The classification result is shown in Table 1.
Table 1 Accuracy of classification result
| Maximum likelihood classification |
59.0% |
| Truncated normal distribution classification |
75.1% |
| Screening ? |
60.9% |
| Screening ? |
62.7% |
| Classification accuracy using both methods |
76.1%. |
As a result of the experiment, classification using both methods obtained the highest classification accuracy compared with independently used method. And, there was seldom undiscriminant pixel. There was no category where the classification accuracy became 10% or less on the each classification category.
6. Conclusion
The screening of training data and maximum likelihood classification using truncated normal distribution, which are techniques of the classification accuracy improvement proposed in this paper, improve classification accuracy. Especially the latter is good method for reducing number of pixels in the undiscriminant class. It also improves the classification accuracy more. It is effective measures for the classification accuracy improvement. And the technique by using both methods improves classification accuracy evidently. In this paper, satellite image of 3 bands of the 30m resolutions was used as object image. If the object image is different from the image used in this experiment, obtained result will be changed. Therefore, we should verify another case, for example different sensor, different time, different resolution and so on.
