Study On Fusion Algorithms And Classincation Methods For GF-1 Data Oriented To Forestland Information

In the past 50 years, remote sensing technology develops rapidly and produces substantial application fruits ranging from small-scale scientific research to large-scale actual production. With the continuous development of 3S technologies, the application of remote sensing technology in forestry monitor industry shows a rapid development and constant innovation on the methods. Based on GF-1 satellite panchromatic and multi-spectral remote sensing data in research area of Liaoning province as the research object, five fusion algorithms are used in the fusion experiment to select the optimal pixel-level image fusion. And on this basis, three different classification methods (artificial visual interpretation and supervised classification, object-oriented classification) are employed to the classification study. Further more, object-oriented classification with multi-scale texture feature is put forward on the basis of object-oriented classification. The research result would provide technical reference for large-scale images fusion and classification application.Content and conclusion of the research are as follows:(l)Focus on the extraction of forestland information, we fused the GF-1 panchromatic and multi-spectral data by using five methods:Principal Component transform、High Pass Fusion、 Brovey transform、Gram-Schmidt transform and Pansharpening. The fusion images were compared and analyzed by using both visual judgments and statistical method. The result indicated that Gram-Schmidt fusion image had a good visual effect, strong ability of spectral preservation, high spectral fidelity and clear texture especially in the forestland. In terms of statistical result of Gram-Schmidt fusion images, the mean was similar to original multispectral images’, the growth of average gradient was more than 10%, each band correlation coefficients with original multispectral band was above 0.700, the high frequency information correlation coefficients of three bands (R, G, B) were as high as 0.900 and NIR band was 0.873. In texture analysis of forestland, Gram-Schmidt fusion images showed highest variance than other fusion methods, the average entropy is 0.729 which increased by 29.0% comparing with original multispectral image, and the second moment presented a 14.9% decrease than original multispectral image. The result showed Gram-Schmidt fusion algorithm is more suitable for GF-1 in image fusion application for forest land information extraction.(2)The traditional classification methods such as artificial visual interpretation and the classification method based on pixels are widely used in forestry industry, and in this paper based on Gram-Schmidt fusion image, artificial visual interpretation, maximum likelihood and minimum distance methods of supervised classification were conducted for the classification experiment. The result showed that the group of five repeating in artificial visual interpretation gained overall accuracy of 0.8065 in average, Kappa coefficient of 0.7677 in average. The classification results of different repeating groups were overall relatively consistent but because of man’s subjective differences the precision showed some fluctuations especially in the internal interpretation of broad-leaved forest and mixed forest land. In terms of supervised classification, the maximum likelihood classification result overall accuracy is 72.21%, the Kappa coefficient is 0.6738, slightly higher than the minimum distance classification of overall accuracy and Kappa coefficient 0.6625,71.52%. The order of forestland classification accuracy is:coniferous forestland> broad-leaved forestland> mixed forestland. Both supervised classification methods produced a large number of fragmentations, and there was an obvious "pepper phenomenon"(3)With the help of ESP (Estimation of Scale Parameters) tool and repeated experiments, we identified the optimal segmentation scale and parameter for the dominant land cover types in study area. Four-level multi-scale segmentation hierarchy (82,63,47,28) was built, and the multi-level classification system was established. The advantages of multi-scale segmentation in object-oriented classification were showed incisively. Gray-level Co-occurrence Matrix was used to extract texture features from the first principal component of PCA, and the best combination of multi-scale was selected by Jeffries-Matusit distance. By analyzing the classification precision of single texture and correlation of textures, best multi-scale textures combination was finally chose to apply to object-oriented classification. The results showed that the method of object-oriented classification with multi-scale texture feature can extract the types of surface features effectively. The precision of classification is 81.75% and Kappa coefficient is 0.78, and compared with object-oriented classification with single-scale texture feature and object-oriented classification with multi-spectral data the precision of classification had increase of 3.48% and 5.99%. In coniferous forest, broad-leaved forest and mixed forest land, compared with single-scale texture method the multi-scale texture method increased by 3.9%,5.9% and 5.9% in precision of classification and gain obvious effects.