A Model Based On Geographically Weighted Regression For Astimating Grass Yield Usinig Remote Sensing Data

The grass yield embodies the grassland productivity closely. And it is important for guiding animal husbandry production management. Nowadays the remote sensing technology has become one important way to estimate the grass yield. There are biological-physical model and experience model. Generally biological-physical model isused for large region in macro scale. Here the remote sensing data with low resolution is always selected as the input parameters. But the accuracy of estimation is low by this means. Experience model can be used to build statistical model between remote sensing data and ground measure data, and it is simple and feasible. While the spatial heterogeneity often be ignored in experience model, which results in a relatively low accuracy. In the paper, the spatial correlation of grass yield was taken into consider creatively. Based the spatial adjacent relationship, the geographically weighted regression (GWR) model was adopted to estimate the grass yield. Here remote sensing data, meteorological data, grassland type data and ground measured data are important input parameters for the GWR.The source region of three rivers (SRTR) was taken as study area in the paper. The correlation ship between ground measured data and other input factors were analyzed firstly, then the factors with high correlation coefficient were used to build the GWR. The ground points were selected carefully to verify the accuracy. In order to illustrate its availability and superiority, another experience model, multiple linear regression model, was selected in comparing. Finally Tibetan Autonomous Prefecture of Golog was chosen as an example, and the grass yield here was estimated with GWR model established and domestic environmental satellite imagery.The main progress and innovation of present paper areas follows:(1) The GWR was adopted to estimate the grass yield with remote sensing data. The result shows the goodness of fit for grass yield estimation model based GWR can be improved significantly. And the fitting value of model r2increases from less than0.3to more than0.8, in the same time the estimation accuracy is much than that of multiple linear regression model, improved by about20%.(2) Several factors, including cumulative precipitation from May to August and drought index from May to August, Modified Soil Adjusted Vegetation Index(MSAVI), have close relationship with grass yield in SRTR. For grass yield estimation model established with these factors, the coefficient of determination r2is0.858, adjusted r20.772, accuracy71.61%. The model proposed in this study is easy to use. Here the input parameters is less, and can be calculated from remote sensing data and meteorological data directly. The model established in the paper can be used to estimate grass yield in SRTR directly.(3) Based GWR, the study has established the technology and methods to estimate regional grass yield using domestic environmental satellite data. And using RS imagery and meteorological data, the regional grass field in TheGolog Tibetan Autonomous Prefecturewas estimated. For the GWR model, r2is0.845, adjusted r20.727, and P less than0.001. The final accuracy of estimation is67.47%. In2010.8, the total yield here (fresh weight) is3260.01×104t. To grass yield there is a gradually decreasing trend from east to west in the region.