Classification of Forest Land Information Using Environment Satellite (HJ-1) Data

For researching properties of HJ-1A CCD camera multi-spectral data in performance on extraction of land features information, this paper selected the east area of NiLeke forest farm in the western Tianshan mountain as the study area, and analyzed different accuracies for HJ-1A CCD data in identifying forest land categories using various classification methods. Firstly, maximum-likelihood classifier, Mahalanobis distance classifier, minimum distance classifier and K-means classifier were used to category land use types with two different scales on HJ-1A CCD1 and Landsat5 TM images, and analyzed separately with confusion matrix. Secondly, forest land types were distinguished by texture information and the smallest polygon size using K-NN method based on clustering algorithm. The comparing results show: at first, different classification system have different accuracy. In the first land use classification system, the accuracy of HJ-1A CCD1 images are lower than TM images, but higher in the second land use classification system. Secondly, accuracy result of maximum-likelihood classification is the best method to classify land use types. In the first land use classification system, TM total accuracy is up to 85.1% and Kappa coefficient is 0.8. In the second land use classification system, the result is up to 85.4% and kappa coefficient is 0.74.Thirdly, judgment both from the view of visual interpretation and quantitative accuracy testes, non-supervised method with K-means classifier has low qualities where many land features have characters of scattered distribution and small different spectrum information. Finally, the experiment proved that there were good vector results on HJ-1A remote sensing image in the view of visual judgment, and extracted deferent forest land by the overall accuracy 87% with the supports by those variables’ distribution knowledge, such as conifer, mixed forest, broadleaf, shrubby.