Research On Key Techniques Of Fast Extraction Of Forest Types From MODIS Image

Although the forest resource monitoring system of our country has been built up, it still exists some problems such as long period, low accuracy and so on. With the fast development of remote sensing and modern information technology such as computer and internet, these questions can be solved effectively. In this paper, three classification algorithms including decision tree, kNN (k-Nearest Neighbor) and method with combination of spectral unmixing were employed to conduct the accurate extraction of forest types (including coniferous forest, broad-leaved forest, mixed forest, bamboo forest, shrub) in China based on the vegetation index products of MODIS data such as NDVI, EVI, LAI, GPP and so on. The major work and results are as follows:(1) A batch preprocessing technology based on the products of MODIS was proposed. With the application of ENVI/IDL programming development language, the technologies of bulk projection transformation, batch image Mosaic, batch removal of black belt and batch image mask were realized within the area of the whole China. Results showed that the data processing effect was better than the traditional image processing methods, especially the processing speed of MODIS image gained a substantial improvement.(2) The decision tree model for forest type recognition was established, which realized the national forest type extraction. An overall classification accuracy of 82.29% with a kappa coefficient of 0.79 was finally obtained for the national average level, met the accuracy requirement of large scale forest information extraction.(3) The kNN algorithm was employed to extract information of forest types for every province of China, and the influence of different k value on the classification accuracy of forest types was compared. It was found that as k value increased, the classification accuracy increased at the beginning, reached its maximum at the k value of 7, and then decreased. Finally an overall classification accuracy of 83.20% with a kappa coefficient of 0.7929 was finally obtained for the national average level, higher than the accuracy of decision tree.(4) As we know there are large mixed pixels in the vegetation index products of MODIS data for the low spatial resolution, and the existence of those pixels will impede the improvement of classification accuracy. Thus the study of forest type cognition with the combination of spectral unmixing was carried out. It was found that with the combination of spectral unmixing, an overall classification accuracy of 84.26% with a kappa coefficient of 0.8022, higher than the accuracy of the other two classification algorithms.Our paper focusing on the problem that it is difficult to achieve fast monitoring of forest resource in large region, proposed meticulous data processing flow for the low resolution vegetation index products of MODIS and feasible technical scheme, which enhanced the accuracy and reliability of national forest type classification. Meanwhile, the spatial distribution map of national forest type provided guidelines and references for the management of regional forest resources.In a word, this paper led to following two innovations:(1) A set of batch preprocessing technology for MODIS data was proposed.(2) A MODIS data based on unmixing technique for the classification of forest types is put forward.