Earth Surface Processes
YANG Xiu-chun, XU Bin, ZHU Xiao-hua, TAO Wei-guo, LIU Tian-ke
There is an ecotone connecting farming region and pasturing region for northern agro-grazing ecotone. Its ecological function consists of conserving water sources, checking the wind and fixing the shifting sand, purifying air and maintaining biodiversity.Grassland is not only one of the important ecosystems, but also a background vegetation. Over the past decades, human activities have caused great land cover changes, such as desertification, grassland degradation, and sandy. Therefore, accurate and timely monitoring grassland is of critical importance for utilizing and administering grassland, developing pasturage and improving ecological environment. Using MODIS remote sensing data for the year 2005 and the ground measured grass yield of the corresponding period, linear regression model,non-linear regression models and BP neural network model were respectively established, to express the regression relationships between ground truth data and vegetation indices in this paper. Some conclusions are drawn as follows: (1) Regional models are better than whole-area general models. It is reasonable for the four grassland areas, and the regional models can better describe grass production.(2) Models based on BP neural network are better than linear regression models and non-linear regression models in fitness accuracy. Its decision coefficient increases by more than 3%, and the highest is 6.92%. Moreover, by precision validating, we find its root mean square error and relative errors are smaller, the models precision increases by more than 2.5%, and the maximum increases 23.22%. It is obvious that models based on BP neural network are most suitable for monitoring grass production of northern agro-grazing ecotone, and it can meet the need of estimating of grass production in northern agro-grazing ecotone.(3) The suitable vegetation indices for monitoring grass production of northern agro-grazing ecotone are NDVI and SAVI.(4) With the accumulation of the temporal scales data, further studies may focus on input data for BP neural network model. For example, input data may adopt soil moisture index and temperature and precipitation, and so on, which may further increase precision of models, and approach actual grass production for monitoring results.