177 
NOAA AVHRR data is relatively low, to obtain higher 
categorized area precision, and it needs to use the pixel 
unmixing method to make the precision of pixel classifications 
reach the level of the sub-pixel. 
Fig.4 Structure of BP Neural Network 
Up to now, the methods of the remote sensing image 
classifications include the traditional supervising classification 
(such as the parallel algorithm, the minimum distance 
algorithm and the maximum likelihood method), 
ISODATA( iterative self-org-anizing data analysis technique) 
not supervising classification, the latest fuzzy classification, 
expert system classification and neural network 
classification(Sun Hong-yu et al,2003). Furthermore, the pixel 
unmixing method includes message extracting methods for sub 
pixel imaging based on linear models, geometrical optics model, 
geometirc model at random, probability model, fuzzy model, 
etc. 
BP neural network is a neural network model adopting the back 
propagation learning algorithm. Nowadays it is applied on 
many fields. The application of images classification and pixel 
unmixing method also make heavy effect(Wang Xi-peng et 
al, 1998). A typical BP neural network is made up of three 
neuron layers, which are the input layer, the output layer and 
the implicating layer (as shown in Fig. 4). 
This text on the basis of NDVI data and the third band AVF1RR 
data with the elevation data of the Yellow River source and the 
temperature precipitation data as the input of BP neural 
network, and the auto-correlation characteristics of 
geographical space and geography experts knowledge and field 
measuring data as the training reference of BP nerve network, 
presents some exercises on the image classification of the 
neural networks and the pixel unmixing function, and 
demonstrates the classifying process of the land cover type of 
the Yellow River source. Among them, the selection of the land 
cover type consults the description of the land cover type of 
Qinghai-Tibet Plateau made by Zhengyi Wu. Finally, we obtain 
the categorized result of the land cover type of the Yellow 
River source in 1990- 2000 years, as shown in Table 2. 
The value region of NDVI reflects the whole trend of area 
changes of land coverage. Also in the table 2 the changes of 
area of land coverage are very distinct. Among which the area 
of all kinds of meadow types show the distinctly decreasing 
trend while the desert type shows increasing tend. However, the 
changes of area of grassland and desert grassland in different 
altitude are different. The area of alpine grassland and the 
representative grassland decreased while the ones of alpine 
desert and alpine grassland, alpine desert, subalpine grassland 
and subalpine desert grassland show the increasing trend. The 
area of desert grassland in lower altitude did not change too 
much. 
Tablet 3 is the correlation modulus between the changes of land 
coverage and the changes of climate factors. The tablet shows 
that there is strong negative relation between the land coverage 
of alpine meadow, alpine grassland, subalpine meadow, 
meadow and grassland and annual average temperature and 
annual highest temperature. While between the area of alpine 
grassland and desert grassland, alpine desert, subalpine desert 
grassland, subalpine desert and desert and the annual average 
temperature and annual highest temperature, there exists a 
strong positive correlation. It suggests that the increasement of 
temperature is a very important inducement of meadow 
degeneration. However, the strong positive correlation between 
annual average temperature and annual highest temperature and 
the area of subalpine grassland indicates the increase of 
temperature accelerates the growth of the vegetation in 
subalpine area. 
The relations between precipitation and the different land 
coverageage type are not accordant. The coverage area of 
meadow shows great negative correlation with precipitation 
which indicates the more precipitation could lessen the value of 
NDVI. While between alpine grassland, subalpine grassland, 
steepe and precipitation, there is a positive correlation which 
suggests the increase of precipitation could improve the 
environment of vegetation growth in normal coverage land. 
Meanwhile, all types of desert area have a weak negative 
correlation with precipitation which indicates that including 
temperature the decrease of precipitation also promote land 
desertification.