Abstract: Abstract: To improve pasture monitoring and low-level digitization in Inner Mongolia Ulanqab desert steppe, the feature extraction and image recognition for two typical pastures were conducted in this paper so as to provide a basis for grass species identification and grassland management. The grass images were collected in the desert steppe of Siziwang banner of Inner Mongolia at 3 pm on August 15, 2014, where belongs to the middle temperate semi-arid continental monsoon climate zone. Under the natural sunlight intensity, 100 original images of varying blade sizes and shapes were taken as training and testing samples from prairie by using an intelligent navigation data acquisition vehicle. The method of color moment was used considering that it can reflect the color distribution information comprehensively. The process doesn't need to quantify the color space, and the feature vector dimension is low when the color feature is extracted. In this process, the first moment describes the mean color, the second moment represents the color variance, and the third moment describes the color skewness. From the three components of Red、Green、Blue (R、G、B), nine color moment feature vectors can be obtained and Hue (H) as well as Saturation (S) components were included along with six color moment feature vectors. But the Value (V) component of the brightness information had nothing to do with the color. In all, 15 dimensional vector features were required to be extracted. Then，RGB and HSV color moment were extracted and color rule set was established for the two typical grasses of Leymus chinensis and Chenopodium glaucum. The results showed that the color moment for these two forages had an obvious difference especially in the first color moment of RGB. The interval range of each parameter was an important factor for image classification and recognition. For the 60 collected images in the steppe, the 2G-B-R vegetation index method was employed separately, for separating the pasture from soil background of which the 33 image separations were successful. Then, the fuzzy C-means clustering algorithm based on 2G-B-R color difference characteristic was applied to the other 27 images of them. All the images acquired ideal background segmentation results. After finishing background segmentation, we built a three-layer BP neural network model. To improve the speed of recognition, the method of principal component analysis was used so that we could reduce the input feature parameters from 15 to 10. The number of output nodes of BP network was 2, which was the number of grass segmentation expressed by the binary form. The number of hidden layer nodes was calculated by the classical formula, the range of values was from 4 to 13, and the experiment indicated that the performance of the network can reach the best condition when the number of hidden nodes was 10 in the hidden layer. By taking 40 images as the validation samples, 89.5% of final overall recognition rate and an effective classification result of two pasture images were received. At the same time, the coverages of 9.78% and 34.21% for Chenopodium glaucum and Leymus chinensis were obtained respectively. The experimental results indicated that the image segmentation and classification were perfectly realized by using Fuzzy C-means clustering algorithm and BP neural network model based on color moment feature. Automatic identification of grass plays an important role in the grass digitization, the segmentation and recognition for forage image provided a scientific basis for the monitoring of vegetation species diversity, grass degradation and pest control. The results of this research could provide a significant theoretical foundation and data support for forage grass species identification. It would provide an important way in realizing the modern grassland ecological environment protection and development of the grassland economy.