Keywords: Landsat, remote sensing image, classification
Chinese Classification Number: P23 Literature Identifier: A Article Number: 1007-3973(2012)007-109-02
1 Analysis of DN value of the image
The reason why remote sensing images can show various features is that each pixel in the image has a different DN value (DN). In order to threshold the remote sensing image, first we need to figure out what position of various features in the remote sensing image, usually we need to figure out the DN value of each feature in the remote sensing image in the interval of the remote sensing image, therefore, we need to analyze the DN value of the remote sensing image first. We choose the Landsat-7ETM+ satellite remote sensing image of Yangshan harbor area in 2010 as the research object, as shown in Figure 1.
This image has the information of all bands of Landsat-7ETM+, and the influence we use is Unsign-8byte image, so the DN value in the image is from 0 to 255, the image is to show different features by the difference of DN value, we can see from the relationship of each DN value of the image Table 1 can show that the remote sensing image of the Yangshan Harbor area in 2010 has 9 bands.
The information extraction method using the threshold relationship method is very easy to realize, we use the relationship between the DN values of each channel, for example, the DN value of the fifth channel is less than 20, we classify it as a body of water, so that the water features can be easily extracted. However, in the classification process, the extraction of various features is not simply dependent on a certain channel to complete.
2 Threshold Classification Algorithm and Classification Model Determination
By analyzing the DN value relationship analysis table in Table 1, we can see that the fifth and eighth channels are very sensitive to water, and the light wave is absorbed by the water, so the DN values are all less than 20, so that we need to pass through the same two channels only to be able to extract the water features, and the vegetation can be extracted through the first and second third channels, and the other features can be extracted through the first and third channels. The second and third channels can be extracted, and several other channels have the phenomenon of overlapping DN values of other features and vegetation features, therefore, the classification can be solved by using ETM+1, 2, 3, 4 to classify vegetation, and human activities have encountered a more difficult problem in this experiment, the main reason is that there is no range of DN values of a certain channel that can be used to extract the features completely, and each channel, when distinguishing between the various features, is not satisfactory. features but unsatisfactory, often water bodies are confused with humans, or vegetation is confused with human activities, in fact, in the course of the experiment, we also found that there is no channel that can extract human activities alone. So we need to introduce a mathematical concept, that is, the concept of intersection, in dealing with human activity features must be intersection analysis, but for the real human activity features is the existence of intersection, we are in the process of threshold analysis, we found that 1, 2, 3, 4, 9 channels, humans and water bodies in the value of greater than 68, 79, 75, 50 and 35, respectively, can be categorized as human and water **** the same kind, and the 5, 8 channels, in DDT, we found that no channel can be extracted from human activities. And 5, 8 channel, in DN values greater than 20 and 20 respectively, can be categorized as human and vegetation *** similar, and the intersection between these two categories is any internal activity features, so we can quickly extract the features of human activities through the concept of intersection. Through the above analysis of threshold algorithm, we need to establish the classification model of this experiment, we select the channels that are useful for classification, and each feature is processed by the intersection in order to get the most matching classification results. The schematic diagram of the classification model is shown in Figure 2.
The method of information extraction using the threshold method is very easy to implement, using only a small amount of human and material resources, can achieve the desired requirements, and can quickly obtain the change information of the region. After integrating the above classification results, we get Figure 3, the classification result map of Yangshan harbor area in 2010.
3 Noise Reduction Processing
The purpose of this experiment is to study and analyze the remote sensing images of the coastal zone of Shanghai, and we need to obtain more accurate classification maps, so that we can provide a good basis for future ecological evaluation calculations, and the noise reduction processing, as one of the three major processes of the remote sensing image preprocessing (radiometric correction, geometric correction, and noise reduction) is also applied in this experiment. Noise reduction process is applied because remote sensing image signals are often affected by various atmospheric effects and ionospheric radiation in the process of generation, transmission, reception and recording, thus generating a variety of noise, which will bring about different impacts when carrying out the next step of remote sensing image processing, such as feature extraction, information analysis, and pattern recognition, so the removal of noise from the remote sensing image before this is a very important preprocessing step. Therefore, remote sensing image noise removal is a very important preprocessing step before this. Most of the pixels in the remote sensing image do not have much difference in gray value, and it is because of this gray correlation that the energy of the remote sensing image is mainly concentrated in the low-frequency region, and only the energy of the detailed part of the image is in the high-frequency region.
The main purpose of remote sensing image smoothing is to eliminate or attenuate the noise on the image, that is, attenuate the high-frequency component and enhance the low-frequency quantity. However, the high-frequency region also contains the detailed energy of the image, so the remote sensing image has a certain attenuation effect on the details of the image while reducing the noise. This process can enhance the low-frequency volume, which means that more complete and coherent geographic features can be obtained when classifying, and such data are of great significance for ecological evaluation. We processed the satellite remote sensing images of Yangshan Harbor area in 2010 by noise reduction, and the images after noise reduction are obviously different from the original images, the image boundaries become smooth and coherent, and the process of extraction of large plaques in the classification process has been significantly improved, as shown in Fig. 3 and Fig. 4, while comparing the classification maps after the noise reduction process and the classification maps without noise reduction, it can be seen that both of them have a higher degree of integrity of the plaques. The image after noise reduction has more advantages, but how the classification effect after noise reduction, we need to evaluate these two images.
4 Classification evaluation
After the DN value analysis, we classify the impact, how effective the classification is, we need a classification standard, we through the combination of a priori knowledge and field measurement data drawn by the standard classification map (Figure 5) as a standard. Describing the classification effect with statistical theory can be done using the linear relationship regression coefficient evaluation method.
According to the mathematical equation of first-order linear regression analysis (1)
Linear regression coefficient
(1)
Through the spatial modeling language function available in Erdas, we programmed and quickly calculated the linear regression coefficient of the classification map 3 with the standard classification r=92.48%. And the linear regression coefficient r=95.19% for the categorization map 4 after noise reduction processing. This indicates that the linear regression of the two images is very good, which shows that the threshold classification approach can improve the classification accuracy of remote sensing images very well. After comparing the noise reduced and un-noise reduced images with the standard map after classification, it is found that the linear regression of the noise reduced image is better than that of the un-noise reduced image and is more consistent with the standard map, which further illustrates that remote sensing images after noise reduction processing are more conducive to the improvement of classification accuracy. This shows that there is a high degree of consistency between our classification map and the standard map, thus indicating that my classification method has a high degree of feasibility.
5 Conclusion
The classification method using threshold analysis can efficiently and accurately classify remote sensing images with high linear regression coefficient r. This paper analyzes and discusses the DN values of various feature characteristics of remote sensing images to find out the relationship between the DN of different feature characteristics, which provides the basis for the classification of the images, and the image classification technique based on DN values can improve the accuracy of classification of remote sensing images in comparison with the traditional remote sensing image classification technique. Compared with the traditional remote sensing image classification technique, the DN value-based image classification technique can improve the efficiency of remote sensing image classification and provide an effective means for remote sensing image information extraction. The classification effect of remote sensing images is better after the noise reduction process. In the process of classification, we also find that the fewer the categories of classification, the easier the remote sensing images can be classified, because the reduction of classification categories can avoid the phenomenon of same object with different spectra, and the phenomenon of same spectra with different objects. Therefore, for the classification of remote sensing images with fewer categories, the classification method based on DN value analysis provides an effective approach for image classification.
(Funded under the Key Project of Shanghai Science and Technology Commission (075105108))
References:
[1] Yin-Xi Li, Xiaoguang Feng, You-Ming Lin. Classification and format of data products of Landsat 7[J]. Remote Sensing Information,2000(3):37-40.
[2] Huang Jianling,Zheng Xuemei. An optimization method for image denoising based on edge detection[J]. Computer Simulation,2009,26(11):260-261.
[3] Jiancheng Luo,Yi Liang,Chenghu Zhou. A hierarchical clustering method based on scale space and its application in remote sensing image classification[J]. Journal of Surveying and Mapping,1999,28(4):319-324.
[4] Xihua Yang. Application of remote sensing technology in the survey of land use status on the Loess Plateau[J]. Remote Sensing of Environment,1990(5):257-265.